PasClaw

PasClaw is an ultra-lightweight personal AI agent written in Delphi Object Pascal. It is a Delphi/FPC port inspired by picoclaw, with a command-line assistant, tool calling, MCP integration, an HTTP gateway, an OpenAI-compatible API surface, a small embedded web UI, scheduled tasks, skills, and channel integrations.

The main program lives at src/pasclaw/PasClaw.dpr. It initializes terminal color handling, prints the banner, applies timezone configuration, and dispatches into the command tree implemented under src/cmd/.

Features

LLM providers — 19-entry catalog covering Anthropic Messages, OpenAI Chat Completions, Google Gemini generateContent, plus OpenAI-protocol-compatible providers (Groq, DeepSeek, Mistral, Cerebras, OpenRouter, Zhipu, Qwen, Moonshot, MiniMax, NVIDIA NIM, Novita, MiMo, VolcEngine, Ollama, vLLM, LiteLLM). The full list lives in src/pkg/providers/PasClaw.Providers.Catalog.pas; adding a provider is a one-record append. Streaming SSE on all three protocol families; native search grounding on Gemini; citation collation on Perplexity. Fallback chain — Cfg.Fallbacks = ["openai", "gemini"] walks the list when the primary returns 429 / 5xx / network error. Code-driven SetProvider('anthropic', $ANTHROPIC_API_KEY) for embedders so no ~/.pasclaw/config.json is required.

Built-in tools — exposed to the model on every turn:

Tool	What it does
fs_read / fs_write / fs_list	sandboxed filesystem access
fs_grep	recursive substring search, returns hashline-formatted matches
fs_edit_hashline	patch by line-anchor + file-hash header, race-safe
shell_exec	/bin/sh -c (or cmd.exe), output capped at 1 MiB, denylist-gated
web_search	DuckDuckGo / Brave / Tavily / SearXNG / Perplexity / Gemini-grounding — 6 providers
web_fetch	HTTP GET → readable plain text (HTML stripped, entities decoded), SSRF-guarded
memory_search	SQLite FTS5 BM25 over workspace/memory/*.md and MEMORY.md (including fetched-*.md written by memory_fetch). When vector_search_enabled is on (default yes), the hybrid backend in src/pkg/memory/PasClaw.Memory.Vector.pas runs a local sentence-transformer embedder (MiniLM by default — embeddings never leave the host) over the same notes, fuses keyword + vector ranks via Reciprocal Rank Fusion, and falls back to FTS-only silently when the runtime artifacts (sqlite-vec extension, ONNX Runtime, model weights) aren't yet provisioned under $PASCLAW_HOME/cache/localvector/.
session_search	SQLite FTS5 BM25 over the full text of every saved session under workspace/sessions/. Searches PAST conversations, not just the current one — returns session id + title + snippet + score with a pasclaw resume <id> hint. Index at workspace/sessions/.search.db, lazily rebuilt (reindexes a session only when its UpdatedAt advances). Gateway stat buckets excluded.
vault_search / vault_get	search + read pasclaw.dev Code Vault entries (Object Pascal samples + components); opt-in via vault_tools_enabled
skill_<name>	Pascal-side tools registered from kind: shell / kind: prompt skills
MCP-bridged	every tool a configured MCP server exports — see below

Parallel dispatch — when the model returns multiple tool_use blocks in one turn, read-only tools (web_search, web_fetch, fs_read/grep/list, memory_search) fan out on worker threads; mutating tools (fs_write, fs_edit_hashline, shell_exec) stay serial. ~50% wall-clock win on multi-network-tool turns.

MCP — both transports. Stdio MCP via spawned subprocess + JSON-RPC over pipes, and Streamable HTTP MCP (handles SSE-framed responses, Bearer-token auth). pasclaw mcp catalog queries the pasclaw.dev MCP registry (GET /api/public/v1/mcp) with a 5 s timeout and falls back to the bundled 5-entry list (replicate, digitalocean-apps, digitalocean-databases, runpod-docs, huggingface) when the hub is unreachable — source attribution (hub / built-in) shown in the output. pasclaw mcp search <q> searches the hub directly (no fallback — bundled list is too small to search). pasclaw mcp install <slug> tries the hub first so any registered server is installable, falls back to the bundled catalog when the hub doesn't have it. Reads the right env var, writes the right Authorization header, never preloaded.

Skills — markdown manifests under $PASCLAW_HOME/workspace/skills/ advertised in the system prompt; the model loads the body via fs_read on demand. Install from GitHub (pasclaw skills install owner/repo[/path][@ref] — codeload zip, FPC's Zipper.TUnZipper or Delphi's System.Zip.TZipFile); from the pasclaw.dev hub (pasclaw skills install hub:<slug>[@<version>], or just pasclaw skills install <slug> which tries pasclaw.dev first then falls back to ClawHub); or from ClawHub directly (pasclaw skills install clawhub:<slug>[@<version>]). pasclaw skills search <q> queries both hubs and aggregates the results (pasclaw.dev first, ClawHub deduped). Malware-flagged skills refused; suspicious-flagged install with a warning.

Code Vault — the pasclaw.dev Code Vault hosts Object Pascal source code (sample programs, reusable components, libraries) as GitHub-backed entries. pasclaw vault search <q> and pasclaw vault show <slug> discover; pasclaw vault install <slug> [<dest>] git clones the entry's repoUrl into $PASCLAW_HOME/workspace/vault/<slug> (or a path you pass). The agent gets two opt-in tools — vault_search and vault_get — registered only when vault_tools_enabled: true in config.json; pasclaw onboard asks during setup with a default-yes prompt. Both tools are read-only HTTP GETs against the vault registry; obtaining the code itself is a separate shell_exec git clone call the model makes after vault_get returns the repoUrl.

HTTP gateway — pasclaw gateway and pasclaw serve. OpenAI-compatible /v1/chat/completions (streaming + non-streaming) and /v1/responses (tool-passthrough so Codex CLI can drive its own tools). Embedded web UI (vanilla ES2020, single HTML file, no JS toolchain) with 8 tabs: Chat (SSE-streamed, tool activity surfaced inline), Memory, Files, MCP, Cron, Skills, Logs (live tail via SSE from the in-process ring buffer), Settings. Read-only inspection endpoints (/v1/mcp, /v1/cron, /v1/skills, /v1/memory, /v1/fs, /v1/config, /v1/logs) backed by the gateway's sandbox + secret-masking.

Chat channels — bidirectional (inbound message → agent loop → outbound reply): Telegram (long-poll bot), Discord (bot polling), LINE (webhook), WhatsApp (Cloud API webhook), Slack (Events API webhook + Incoming Webhook reply), Matrix (REST /sync long-poll, federated, self-hostable), IRC (TIdIRC), Email (SMTP send + IMAP poll, --email flag, env-var configured). Outbound-only: Microsoft Teams (Incoming Webhook), generic Webhook sink.

Cron — pasclaw cron add daily-summary "0 9 * * *" summarize "workspace/memory". Last-fired timestamp persisted to workspace/cron/state.json so a missed slot (gateway down, laptop closed) catches up on the next tick instead of being silently skipped. At-least-once delivery — the state file is written after the skill runs, so a crash in the window between the skill's side effects and the timestamp persist will replay the job on restart; idempotent skills are safe, side-effecting skills should self-deduplicate. Per-entry channel sink (--channel <kind>:<target>) posts skill output; output also appended to workspace/memory/<today>.md for the model to recall later.

Memory — SQLite FTS5 BM25 index over workspace/memory/*.md and MEMORY.md. pasclaw migrate (re-)indexes; pasclaw membench --records N benchmarks. Conversation history compaction (pasclaw compaction.threshold_tokens) kicks in mid-loop, summarises the older portion via the same provider, folds the summary into the system prompt, falls back to verbatim on summariser failure (no silent context loss). memory_fetch(url, name?) (registered when web_fetch_enabled: true) fetches a URL and writes it to workspace/memory/fetched-<sanitised>.md with a 4-line provenance header — the body never enters context; the next memory_search indexes it via SyncDir. Auto-dedup: a second memory_fetch against the same URL within 24h short-circuits the HTTP and returns "already indexed (cached Nh ago)" — the existing cached file's source: and fetched_at: header lines are matched against the request before any network round-trip. Inspired by chopratejas/headroom's cross-agent memory dedup.

Working-state snapshot — each session's meta.working_state records the last 8 file paths edited via fs_write/fs_edit_hashline, the most recent shell_exec command, and the most recent tool-call error. Persisted under the session JSON's meta.working_state object; rebuilt after every successful tool loop and re-injected as a system-prompt prefix on the next turn. Survives /quit-and-resume, compaction, and pasclaw agent --session <id> so the agent picks up with structured edit/shell/error context even when the conversation transcript no longer carries it.

Sandbox + safety — read/write path allowlists, shell-command denylist (separate restrict_to_workspace denylist + shell_deny_enabled global), SSRF guard on web_fetch (IPv4 blocklist incl. 169.254.169.254, redirect re-check), hashline patches require matching file-hash header (stale patches abort without writing). TLS required for HTTPS provider/MCP calls via Indy's OpenSSL IO handler.

Embedding — TPasClawAgent and TPasClawServer as TComponents in PasClaw.Agent:

uses PasClaw.Agent, PasClaw.Tools;

Agent := TPasClawAgent.Create('claude-opus-4-7');
Agent.SetProvider('anthropic', GetEnvironmentVariable('ANTHROPIC_API_KEY'));
Agent.RegisterTool(TWebSearchTool.Create);
Agent.RegisterTool(TFileSystemTool.Create);
WriteLn(Agent.Run('Summarize the latest Delphi release notes.'));
Agent.Free;

Form-designable with published properties for the VCL/FMX path; code-driven OOP API for everything else. Custom tools subclass TPasClawTool and override Name / Description / Schema / Run / Category. Single-process server: TPasClawServer.Create('0.0.0.0', 8088); Server.Run; blocks until Stop is signalled from another thread.

Interactive chat — pasclaw agent and pasclaw tui ship slash commands: /help lists them; /status shows model + provider + message count + thinking state; /new starts a fresh session (new id, history cleared); /reset clears history in the current session; /compact forces a summariser pass; /think toggles extended-thinking mode for the next turn (Anthropic Claude); /tools lists registered tools; /quit exits.

Persistent sessions — pasclaw agent serialises conversation history to $PASCLAW_HOME/workspace/sessions/<id>.json after every turn (messages + tool_calls + tool_results + model + provider + compaction summary). Persistence is on by default — every interactive pasclaw agent run auto-allocates a fresh id and the conversation survives Ctrl-C / crash without any flag. Resume with pasclaw resume <id> or pasclaw agent --session <id>; an id that doesn't yet exist starts a fresh session at that id (handy for scripts pre-seeding e.g. daily-2026-06-01). pasclaw session list / show / delete / export manages saved sessions. Session ids are yyyymmddTHHMMSS-<8 hex> — sortable and collision-safe. On the Delphi build of pasclaw tui the same persistence is wired through a positioned two-pane TUI (session list on the left, chat on the right) — see Chat and UI below. The FPC build of tui keeps history in-memory.

Prompt caching — on by default for Anthropic and OpenAI. The Anthropic request builder emits cache_control: { type: "ephemeral" } on the system prompt and the trailing tools-array entry (uses 2 of the 4-breakpoint budget; caches the largest stable prefix on every turn), with optional ttl: "1h" extended-TTL hint. OpenAI gets prompt_cache_key anchored to the persistent session id so each conversation routes to its own cache bucket. Cache hit / write tokens roll up in /status (cache: on, N read / M written (X% hit on input so far)) and surface inline in the per-turn [tokens in=… out=… cache_r=… cache_w=…] summary. Disable via prompt_cache.enabled: false in config.json; extend the TTL via prompt_cache.ttl: "1h". Gemini caching not yet wired — implicit caching applies automatically on supported models.

Sender identity — every channel tags inbound messages with a canonical <platform>:<id> (e.g. slack:U12345, telegram:5551234, matrix:@eli:matrix.org, email:[email protected], irc:eli, discord:9876543210, cli:$USER). Identity rides on TToolLoopConfig.Identity from the channel boundary down through hooks and audit logs, so embedder hooks can gate behaviour per-sender. Configure an allowlist in config.json:

"allow_senders": ["slack:U-eli", "telegram:*", "cli:*"]

Patterns are exact ids or <platform>:* wildcards; * allows anyone (escape hatch). Empty array (default) = no gate. Each channel calls IsAllowedSender before invoking the agent — non-matching senders are dropped at the boundary with a log line, the model never sees them. Ports picoclaw's pkg/identity pattern.

Mid-loop steering — push a follow-up into a running agent without waiting for the current tool loop to finish. From another terminal:

pasclaw steer 20260601T093015-1a2b3c4d "actually skip X, focus on Y"
pasclaw steer <session-id> --list        # show pending count
pasclaw steer <session-id> --clear       # drop the queue

The running pasclaw agent --session <id> drains the queue at the top of its NEXT tool-loop iteration and folds each pending message into history as a [user steering] ... system note before the next LLM round-trip. Up to 4 messages per iteration are applied (MaxSteeringPerTurn, matching nanobot's _MAX_INJECTIONS_PER_TURN); extras are dropped with a warning so a runaway pusher can't grow history unbounded. /steer <msg> inside an interactive session is the same mechanism (handy for testing). /reset, /new, and pasclaw session delete clear the queue. Storage is one append-only JSONL file per session under $PASCLAW_HOME/workspace/steering/<id>.jsonl — atomic per-line POSIX appends + rename-to-tmp-on-drain so concurrent push/drain doesn't lose messages. Currently only CLI sessions wire SteeringKey; channels can opt in by passing their own per-conversation key on TToolLoopConfig.SteeringKey once they restructure to non-blocking poll loops.

Hooks + steering — TPasClawHook (in PasClaw.Agent.Hooks) is the typed callback surface for embedders to observe, transform, or veto agent events. Four virtuals: BeforeTurn(var ContinueTurn, var Messages) — set ContinueTurn := False to abort cleanly; BeforeToolCall(call, var Cancel, var SyntheticResult) — Cancel := True bypasses the real tool handler and uses SyntheticResult as the tool_result (the approval-gate pattern); AfterToolResult(call, var ResultText, var ErrMsg, var SteeringMessage) — rewrite results inline AND inject a system note before the next LLM round (picoclaw's steering); OnError(Stage, Msg) — observe failures. Register via Agent.RegisterHook(THook.Create); multiple hooks form an ordered chain in registration order.

Subagents — fan-out to focused specialists via a spawn(agent, prompt) tool. Declare them in config.json's subagents: array (name + description + system prompt + tool allowlist + optional model / max-iterations override). Each spawn runs a short RunToolLoop against the parent's provider + fallback chain with a registry filtered to the named tools and a specialist system prompt; result lands back as the parent's tool_result. Implementation in src/pkg/agent/PasClaw.Agent.Subagent.pas — picoclaw's SubTurn pattern, nanobot's subagent module, openclaw's multi-agent routing, ~300 LOC.

"subagents": [
  { "name": "researcher",
    "description": "Web search + summary specialist",
    "system_prompt": "You search the web and produce a 3-bullet summary...",
    "tools": ["web_search", "web_fetch"],
    "max_iterations": 4 },
  { "name": "coder",
    "description": "Code editor",
    "system_prompt": "You edit code precisely using hashline patches...",
    "tools": ["fs_read", "fs_write", "fs_grep", "fs_edit_hashline"] }
]

Cross-platform — Linux x86_64 + aarch64 under FPC 3.2+; macOS x86_64 + arm64 under FPC (Homebrew unit paths autodetected); Windows x64 + Linux + macOS under Delphi 12 / RAD Studio. Windows-on-ARM64 builds via FPC are supported through the Makefile's CROSS_TARGET=aarch64-win64 override (pair with FPC_UNITS_DIR pointing at the cross-build's unit tree); the updater emits windows_arm64.exe as the release-asset suffix on those builds. The Delphi 13 WinArm64EC target isn't wired into PasClaw.dproj yet — add the platform via the IDE's Project Manager when you're ready to ship for it. Three sample binaries under samples/component-console/ plus matching .dproj files for RAD Studio and dcc32.cfg / dcc64.cfg for cmdline Delphi builds.

Changelog

Notable feature additions, newest first. Bug fixes and review follow-ups are in the full git log.

• 2026-06-10 — TUI: subagents wired in (spawn + the background quartet register when config.json declares subagents — previously CLI-agent-only) and the /stats + /model overlays no longer flicker over the session pane (the every-50ms repaint painted pane content then the overlay box on top each tick; panes now freeze while these static modals are up; the /theme menu keeps live repaint for its preview).
• 2026-06-10 — Background subagents (Claude Code shape): spawn_background(agent, prompt) returns a handle and the parent loop keeps working; when the job finishes its result is pushed into the next loop iteration as a [background subagent results] block in the system prompt (same channel steering and compaction use). Companion tools: spawn_status(handle), spawn_wait(handle, timeout_sec), spawn_cancel(handle). Up to 4 concurrent jobs per session; each runs RunToolLoop against a filtered child registry on its own thread. Jobs die with the session — coordinator teardown abandons still-running workers and they self-free. Synchronous spawn from PR #93 is unchanged.
• 2026-06-10 — session_search tool: FTS5 keyword search over the full text of every SAVED session, not just the current one (hermes-agent's session_search). Index lives at workspace/sessions/.search.db, lazily rebuilt from the JSON transcripts (reindexes a session only when its UpdatedAt advances). Returns session id + title + snippet + BM25 score with a pasclaw resume <id> hint per hit. Gateway stat buckets are excluded. Reuses memory_search's FTS query normaliser so both tools treat a query identically. (Fetched-doc search — the other half of the original gap — was already covered: memory_fetch writes into workspace/memory/ which memory_search already indexes.)
• 2026-06-10 — Tool-RPC callback from execute_code: inside a bash/PowerShell script the model can call back into the same tool registry it's been given via pasclaw __tool <name> '<json-args>'. Replaces N + 1 + N inference rounds (list → loop fs_read each → summarise) with a single execute_code whose script body calls memory_search / fs_read / web_fetch as many times as it likes. Loopback TCP server on a kernel-allocated port + token authentication; connection info written to $PASCLAW_HOME/run/tool-rpc.json.
• 2026-06-09 — Gateway stats buckets: stateless /v1/chat, /v1/chat/completions, and /v1/responses calls now accumulate per-endpoint into synthetic session files (_gateway_v1_chat_completions.json etc.) so the web UI Stats tab actually ticks up when the operator uses the web chat (previously gateway requests had no session and showed zero). Each endpoint gets one bucket; concurrent calls serialised through a single critical section.
• 2026-06-09 — Auto-router (UltraCode-Shim shape): heuristic classifies each user message as easy/abstain/hard from keyword markers + token count + tool mix, and routes "easy" turns to a cheap-tier fallback provider while leaving real work on the primary. Opt-in via auto_router.enabled (onboarding asks); the primary is automatically prepended to the fallback chain on routed turns so a misclassification falls through cleanly.
• 2026-06-09 — Per-session stats persistence + /v1/stats endpoint + web UI Stats tab. Opt-in via stats_collection_enabled (default off; onboarding prompts). When on, each turn's token / tool-call / truncation-savings counters accumulate into the session JSON; the gateway exposes a 5-second-cached aggregate at GET /v1/stats with by_provider and by_model rollups; the web UI renders an auto-refreshing Stats tab. TUI /stats overlay is unchanged (it keeps its own in-process accumulator and works regardless of the flag).
• 2026-06-09 — execute_code tool: model writes a multi-line bash or PowerShell script, body is materialised to a temp file under $PASCLAW_HOME/tmp/ and run by the right shell. Cuts inference rounds on fan-out / loop / heredoc patterns that would otherwise need N separate shell_exec calls. Same sandbox + denylist gate as shell_exec.
• 2026-06-09 — pasclaw runbook asks the agent to probe the current project and write a starter ./AGENTS.md (project overview, build / test commands, repo layout, conventions, gotchas). Refuses to clobber an existing AGENTS.md without --force; pairs with execute_code so the probe is one tool call instead of a dozen.
• 2026-06-09 — pasclaw learn --write-scars emits / refreshes workspace/memory/SCARS.md with Atlas-style stable §ANCHOR-NAME ids per recurring failure (citable from commits and PRs via git log --grep "§"); operator-edited Root-cause / Do / Do-NOT rationale survives re-runs.
• 2026-06-09 — pasclaw export [agents|claude|cursor|gemini|zed|all] renders the operator's MEMORY.md + SCARS.md + skill manifests + sandbox policy into the rules files Claude Code (CLAUDE.md), Cursor (.cursor/rules/agent.mdc), Gemini CLI (GEMINI.md), Zed AI (.zed/agent.md), and the cross-runtime AGENTS.md convention already look for at the repo root, so teammates using a different agent runtime pick up the same project rules without installing PasClaw.
• 2026-06-08 — Server-side call_id → thoughtSignature cache on /v1/responses so stock OpenAI-Responses clients (Codex CLI etc.) round-trip Gemini 3's signed function calls across turns even when they strip PasClaw's provider_signature extension. (#194)
• 2026-06-08 — memory_fetch URL auto-dedup: a second fetch against the same URL inside 24h short-circuits the HTTP and returns the cached path. (#192)
• 2026-06-08 — pasclaw learn mines session transcripts for recurring tool failures, normalises into clustering signatures (paths/pids/hashes stripped), reports patterns above --min-sessions, optionally appends a dated block to MEMORY.md. (#190)
• 2026-06-08 — Per-command shell output filters with tee-on-failure: git status/diff/log, pytest/cargo test/npm test, grep/Select-String/findstr, ls -R/Get-ChildItem/dir /s. Cross-shell dispatch normalises PowerShell / cmd.exe / shell wrappers. Counters surface in TUI /stats. (#189)
• 2026-06-08 — In-browser wasm build via container2wasm: make C2W=1 browser produces a single-page agent UI with fetch-proxy networking, onboarding and chat run inline in the tab. (#185, #186, #187, #188)
• 2026-06-08 — TUI session pane sorted newest-first by UpdatedAt, assistant markdown rendered as ANSI in the chat pane, pasclaw tui resumes the newest session on launch, no sort drift on session-pane navigation. (#182, #184)
• 2026-06-08 — Context-mode follow-ups: memory_fetch tool (writes URL bodies to workspace/memory/fetched-*.md, body never enters context), per-session working-state snapshot (recent edits / shell / errors) re-injected as system-prompt prefix after compaction, think-in-code steering rule. (#180)
• 2026-06-08 — Per-tool-result output truncation cache (tool_output_cap config) with tool_output_get(handle, offset, len) retrieval; TUI /stats overlay shows token totals, truncation savings, per-tool call counts. (#176)
• 2026-06-07 — Live /v1/models discovery with on-disk cache keyed on operator-facing provider name, onboarding picker, TUI /model modal switcher (Up/Down/Enter, FPC line-based fallback) with inline auto-refresh when the cache is empty. (#171, #173, #175)
• 2026-06-06 — Hybrid FTS5 + local-vector memory_search backend with in-tree localvector port and a pasclaw memory provision command that fetches sqlite-vec, ONNX Runtime, and the MiniLM embedder. (#165, #166)
• 2026-06-06 — Catalog adds xAI (Grok) + LM Studio entries; Moonshot default bumped to kimi-k2.6. (#163)
• 2026-06-06 — vector_search_enabled config flag + onboarding question for the hybrid memory backend. (#164)
• 2026-06-05 — Gemini server-side google_search grounding, default-on with Gemini-3-or-later gating for the function-calling combo. (#158)
• 2026-06-05 — Gemini 3 thoughtSignature round-trip on functionCall parts for tool-use continuations. (#154)
• 2026-06-05 — Markdown rendered as ANSI-styled text in the terminal for agent + TUI surfaces. (#155)
• 2026-06-05 — OpenAI server-side web_search_options support, default-on for genuine OpenAI endpoints only. (#146)
• 2026-06-05 — web_fetch tool gated behind explicit web_fetch_enabled opt-in. (#144)
• 2026-06-04 — pasclaw.dev Code Vault tools (vault_search / vault_get) for Object Pascal sample discovery. (#130)
• 2026-06-04 — MCP catalog + hub with one-command built-in server installs. (#131, #134)
• 2026-06-04 — Onboarding prompt for built-in MCP servers. (#126)
• 2026-06-01 — TUI themes, widgets, and chat layout polish. (#122, #123, #124)
• 2026-06-01 — Docker image and Dockerfile for containerised deployment. (#121)
• 2026-06-01 — Persistent sessions, resume, and mid-loop steer follow-ups. (#117, #120)
• 2026-06-01 — Anthropic prompt-caching toggle with TTL config. (#118)
• 2026-06-01 — Channel sender identity + allowlist gating. (#119)
• 2026-06-01 — Windows ARM64 cross-build target. (#115)
• 2026-06-01 — Agent hooks and steering API. (#113)
• 2026-05-30 — Tier-1 features: parallel tool dispatch, code-driven provider config, Delphi sample projects. (#98, #96, #95, #101)
• 2026-05-30 — MCP catalog of built-in servers. (#91)
• 2026-05-30 — Web UI with chat / sessions / memory / MCP / skills / vault tabs. (#88)
• 2026-05-29 — Channels: Matrix + IRC + WhatsApp + LINE bots in addition to Telegram. (#75, #76, #77, #86)
• 2026-05-29 — Web search across DuckDuckGo / Brave / Tavily / SearXNG / Perplexity / Gemini. (#80, #81, #82)
• 2026-05-29 — SSRF guard blocking cloud-metadata + RFC1918 + loopback in web_fetch. (#85)
• 2026-05-29 — Workspace sandbox with cd-gating and shell denylist. (#84)
• 2026-05-29 — Cron persistent state + per-job channel sinks. (#79)
• 2026-05-29 — FTS5-backed memory_search tool. (#74)
• 2026-05-28 — OpenAI Responses API on the gateway. (#62)
• 2026-05-28 — Skills system: SKILL.md manifests, GitHub install, ClawHub install. (#55, #56, #59)
• 2026-05-28 — Gemini provider (generateContent REST). (#48)
• 2026-05-27 — pasclaw serve — OpenAI-compatible HTTP server with SSE streaming. (#20)
• 2026-05-27 — Provider catalog (19+ kinds: Anthropic, OpenAI, Groq, OpenRouter, Mistral, ...). (#40)
• 2026-05-27 — Delphi visual component (TPasClawAgent). (#37)
• 2026-05-27 — Hashline diff format port for fs_edit_hashline. (#26)
• 2026-05-27 — TUI (full-screen interactive console). (#24)
• 2026-05-26 — Initial Delphi/FPC port of picoclaw. (#2)

Requirements

• Free Pascal 3.2+ in Delphi mode, or Delphi/RAD Studio.
• Indy (TIdHTTP, TIdHTTPServer) for HTTP clients, the gateway, and channel integrations.
  • FPC builds vendor Indy into vendor/Indy.
  • Delphi/RAD Studio ships Indy, so no vendored Indy checkout is required.

Build

Delphi / RAD Studio

Open src/pasclaw/PasClaw.dproj in Delphi/RAD Studio and build the project. The checked-in Delphi project already contains the project search paths.

On Windows, you can optionally build from a RAD Studio command prompt with:

build-delphi.bat

The batch file uses MSBuild with the existing Delphi project when available, and falls back to the installed Delphi command-line compiler.

Free Pascal

See fpc.md for detailed FPC prerequisites, Indy vendoring, resource generation, Makefile targets, and variable overrides.

On Windows, you can optionally build with:

build-fpc.bat

Configuration

PasClaw stores configuration as JSON. By default:

• Home directory: ~/.pasclaw
• Config file: ~/.pasclaw/config.json
• Default provider: anthropic
• Default model: claude-opus-4-7
• Gateway bind address: 127.0.0.1
• Gateway port: 8088
• Gateway log level: info

Environment variables:

Variable	Purpose
PASCLAW_HOME	Overrides the PasClaw home directory.
PASCLAW_CONFIG	Overrides the config file path.
PASCLAW_VERSION	Compile-time FPC version override used by the Makefile.
PASCLAW_TELEGRAM_TOKEN	Default Telegram bot token for pasclaw gateway --telegram.
PASCLAW_LINE_TOKEN	LINE Messaging API channel access token. Used by pasclaw post line and pasclaw gateway --line.
PASCLAW_LINE_SECRET	LINE channel secret. Required by pasclaw gateway --line to verify X-Line-Signature on inbound events.
PASCLAW_WHATSAPP_TOKEN	WhatsApp Cloud API system-user access token. Used by pasclaw post whatsapp and pasclaw gateway --whatsapp.
PASCLAW_WHATSAPP_PHONE_ID	WhatsApp phone-number ID (the numeric ID, not the phone number itself).
PASCLAW_WHATSAPP_VERIFY_TOKEN	User-chosen string used to verify Meta's GET /webhooks/whatsapp subscription handshake.
PASCLAW_WHATSAPP_APP_SECRET	Meta App Secret used to validate X-Hub-Signature-256 on inbound events.
PASCLAW_BRAVE_API_KEY	Brave Search API key for the web_search tool when web_search.provider = brave.
PASCLAW_TAVILY_API_KEY	Tavily API key for the web_search tool when web_search.provider = tavily.
PASCLAW_SEARXNG_API_KEY	Bearer token for protected SearXNG instances (most public ones don't need it).
PASCLAW_PERPLEXITY_API_KEY	Perplexity API key for the web_search tool when web_search.provider = perplexity.
PASCLAW_GEMINI_API_KEY	Google AI Studio key for the web_search tool when web_search.provider = gemini. PASCLAW_GOOGLE_API_KEY works too.
PASCLAW_MATRIX_HOMESERVER	Matrix homeserver base URL (e.g. https://matrix.org) for pasclaw gateway --matrix.
PASCLAW_MATRIX_TOKEN	Matrix access token (provisioned out-of-band via /login or the homeserver admin UI).
PASCLAW_IRC_SERVER	IRC server hostname (e.g. irc.libera.chat) for pasclaw gateway --irc.
PASCLAW_IRC_PORT	IRC server port (default 6667).
PASCLAW_IRC_NICK	IRC nickname the bot connects with.
PASCLAW_IRC_CHANNEL	IRC channel to join on connect (must start with #).
PASCLAW_IRC_PASSWORD	Optional NickServ / server password.
NO_COLOR	Disables ANSI color output.

Useful config commands:

pasclaw onboard       # create/update home + workspace dirs, pick a provider + model + API key,
                      # then optionally enable any of the 5 built-in MCP servers inline
                      # (replicate, digitalocean-apps, digitalocean-databases, runpod-docs, huggingface)
pasclaw config        # print current JSON config
pasclaw config path   # print resolved config path
pasclaw config reset  # write a default config

Command surface

Global flags:

pasclaw --help
pasclaw --no-color status
NO_COLOR=1 pasclaw status

Top-level commands are dispatched by src/cmd/PasClaw.Cmd.Root.pas:

Command	Purpose
config	View or reset the JSON configuration.
onboard	Initialize PASCLAW_HOME, workspace folders, and provider settings.
agent	Chat with the assistant from the terminal.
tui	Chat in the full-screen TUI.
auth	Store, clear, or inspect provider API keys.
gateway	Start the full HTTP gateway, embedded web UI, cron scheduler, tools, MCP, and optional Telegram channel.
serve	Start the OpenAI-compatible API server surface.
status	Show provider, model, gateway, MCP, cron, and skill status.
cron	Manage scheduled tasks.
mcp	Manage MCP server entries.
migrate	Run data migrations for older versions.
skills	List, install, or remove skill extensions.
model	Show or change the default model.
post	Send a one-shot message to a Discord, Slack, Microsoft Teams, generic, LINE, or WhatsApp webhook target.
membench	Benchmark the memory log subsystem.
update	Check GitHub releases or self-update.
version	Print version/build information.

Chat and UI

pasclaw agent
pasclaw agent -m "hello"
pasclaw agent --model claude-opus-4-7 --provider anthropic -m "summarize this repo"
pasclaw agent --system "Be concise" --thinking medium --max-tokens 2048 --max-iterations 25
pasclaw agent --no-tools --no-mcp --no-hashline

pasclaw tui
pasclaw tui --session 20260601T093015-1a2b3c4d
pasclaw tui --provider openai --model gpt-4o-mini
pasclaw tui --theme matrix
pasclaw tui --no-tools --no-mcp --no-hashline

The Delphi build of pasclaw tui paints a positioned two-pane layout (MVCFramework.Console-backed, default theme ConsoleThemeDefault): session list on the left, chat history + input on the right, header bar with provider/model/clock, footer bar with focus-aware hotkeys. Tab swaps focus between panes; in the session pane Up/Down navigates, Enter loads, N spawns a fresh session, D then Y deletes (one-key confirm), R refreshes the list. In the chat pane typing fills the input; Enter sends the turn (which runs in a background thread so the UI keeps redrawing — spinner shown in the divider above the input); Up/Down scrolls the chat back. If a follow-up is sent while a turn is still running it's queued via pasclaw steer automatically. Pending steering count is shown in the chat divider next to the spinner. Q or Escape quits. The FPC build of pasclaw tui stays on the original line-based ANSI renderer (no session list pane yet).

Theme switcher — --theme <name> picks the initial colour theme; type /theme inside the TUI to open a modal menu with Up/Down live-preview, Enter to apply, Escape to revert. Themes are the ones DMVCFramework ships: default (default), navy, matrix, sunset, ocean, midnight, classic. Unknown names fall back to default silently. Changes don't persist across runs — pass --theme next time, or wire a config field in a follow-up.

Model switcher — type /model inside the TUI to open a modal picker of the cached /v1/models roster for the session's provider (falls back to Cfg.DefaultProvider). Up/Down navigates, Enter applies, Escape cancels. The picker reads the cache file populated by pasclaw model refresh <provider> / pasclaw onboard, and the modal's subtitle shows how stale that cache is. Picking a model updates the active session's meta.model (persisted on the next turn's auto-save) but does not touch config.json — use pasclaw model set for that. When the cache is empty or missing, the TUI auto-refreshes inline: it flashes fetching models from <provider>..., runs /v1/models on a background thread (so the UI stays responsive), writes the cache, and opens the picker when the worker completes. Failures flash an error and leave you on the existing screen. The 15-second wall clock guards against a stuck HTTP. The FPC build exposes /model as a line-based command with the same cache-first → fetch-fallback behavior, but synchronous (line-based UI doesn't fight a render thread): bare /model prints the cached list and current selection (fetching live if no cache exists), /model <id> switches the active model for the running TUI.

Tool output truncation — large tool outputs (recursive greps, multi-thousand-line file reads, web fetches) routinely dump 20-50 KB of raw bytes into the context window; after a dozen of those, the bulk of the budget is tool noise the model rarely needs in full. Set tool_output_cap in config.json to a byte cap (e.g. 8192 ≈ 2K tokens) and RunToolLoop diverts oversize tool results into a process-lifetime cache, replacing the in-context body with [tool output truncated: N bytes, handle=...] plus a head/tail snippet. The model dereferences via the auto-registered tool_output_get(handle, offset, len) tool when it actually needs the rest. 0 (default) is off — legacy verbatim behavior. Errors stay verbatim regardless of cap (they're short and the head/tail split would just obscure the failure).

Shell-output filters — shell_exec is the loudest tool: git status, cargo test, pytest, npm test, recursive greps, and recursive listings all spew kilobytes of mostly-noise on the success path. PasClaw.Tools.Shell.Filters runs each successful shell return through a per-command condenser that keeps the signal (failure traces, counts, the first few changed paths) and drops the rest. The dispatcher canonicalises PowerShell aliases (Get-ChildItem/gci/dir → ls, Select-String/sls/findstr → grep, Get-Content/gc/type → cat) and peels common shell wrappers (powershell -Command "...", cmd /c ..., bash -c "...") so the same filter fires whether the model is talking to Linux/macOS or Windows. Tee-on-failure: when the command exits non-zero, the filter is bypassed entirely — the model gets the full output for debugging. Unknown commands always pass through (no surprises for one-off invocations). Filtered byte savings surface in the TUI /stats overlay alongside the OutputCache numbers. Inspired by rtk-ai/rtk.

pasclaw learn — session-failure mining — walks $PASCLAW_HOME/workspace/sessions/*.json and scans every mrTool message for failure-shaped lines (error:, command not found, permission denied, no such file, non-zero exit=, etc.). Each candidate normalises into a clustering signature that strips per-session variability — digit runs ≥ 2 collapse to <n>, hex runs ≥ 8 (with at least one letter) to <hash>, absolute paths to <path> — so "fs_write to /tmp/abc123/foo.pas line 42 failed" and "fs_write to /tmp/xyz789/foo.pas line 17 failed" cluster as one pattern. Patterns above --min (default 2 occurrences) print sorted by frequency with the tool that triggered them, a verbatim sample, and a first/last-seen window. --write appends a dated ### Patterns observed block to workspace/memory/MEMORY.md so the next agent loop picks it up via the system prompt. --write-scars additionally emits workspace/memory/SCARS.md with Atlas-style stable §ANCHOR-NAME ids per pattern — citable from commit messages and PRs via git log --grep "§" so a fix links back to the failure it addresses; operator-edited Root-cause / Do / Do-NOT rationale survives re-runs (only new patterns get fresh anchor blocks). --since <days> limits the scan window so old sessions don't keep resurfacing fixes the operator has already applied. Inspired by chopratejas/headroom's headroom learn (mining + MEMORY.md) and Abbasi-Alain/atlas (SCARS anchor convention).

pasclaw export — cross-runtime adapter — renders PasClaw's MEMORY.md, SCARS.md, skill manifests, and sandbox policy into the rules files other agent runtimes already look for at the repo root, so a teammate using Claude Code, Cursor, Gemini CLI, Zed AI, or any AGENTS.md-aware tool gets the same project context without installing PasClaw. pasclaw export (default all) writes ./AGENTS.md, ./CLAUDE.md, ./.cursor/rules/agent.mdc (with the YAML alwaysApply: true frontmatter Cursor requires), ./GEMINI.md, and ./.zed/agent.md; single-target form (pasclaw export claude) writes just one; --stdout prints the rendered body for piping; --to <dir> targets a directory other than CWD. Each generated file carries a <!-- generated by: pasclaw export --> marker on the first line — re-runs clobber, hand-edits belong in the source under $PASCLAW_HOME/workspace/. Inspired by Abbasi-Alain/atlas's sidecar-export approach.

pasclaw runbook — bootstrap AGENTS.md — one-shot agent invocation that probes the project (top-level layout, Makefile / package.json / Cargo.toml / etc. for build + test commands, recent git history) and writes a starter ./AGENTS.md covering project overview, build, test, repo layout, conventions, and gotchas. Refuses to overwrite an existing AGENTS.md without --force; --to <file> targets a different filename. Pairs with the new execute_code tool so the probe is one tool call (the model writes a multi-line bash script) instead of a dozen separate shell_exec rounds. Use after pasclaw onboard on a fresh repo, before pasclaw export curates the file with PasClaw-internal state.

execute_code tool — model writes a multi-line bash or PowerShell script body via execute_code({lang, code}); PasClaw materialises the body to a temp file under $PASCLAW_HOME/tmp/exec-*.{sh,ps1}, spawns the right shell (bash <file> on unix, pwsh -NoProfile -ExecutionPolicy Bypass -File <file> on Windows), captures combined stdout+stderr, and cleans up the file. Borrowed conceptually from picoclaw's Python execute_code (bash/powershell instead of Python because both already ship with the host OS — no extra install). The same PasClaw.Tools.Sandbox.ShellAllowed denylist that gates shell_exec runs against the script body, so a buried rm -rf / inside a heredoc gets refused identically to the same string passed inline. Cwd is pinned to the workspace when sandbox.restrict_to_workspace is on, same as shell_exec. The big win: fan-out patterns (list every *.pas, grep each for a symbol, summarise hits) that would take 1+N+1 inference rounds under shell_exec (because the model can't comfortably escape a for loop into a single-line command) collapse to one tool call. Output runs through the tool_output_cap truncation cache, same as every other tool. Inside the script the model can also call back into the same tool registry via pasclaw __tool <name> '<json-args>' — see the next paragraph.

Tool-RPC callback (pasclaw __tool) — when execute_code spawns a script, PasClaw starts a loopback TCP server on a kernel-allocated port and writes the port + a fresh 32-hex-char token to $PASCLAW_HOME/run/tool-rpc.json. The hidden pasclaw __tool <name> '<json-args>' subcommand (not in --help) reads that file, opens TCP, sends {"token":"...","name":"...","args":{...}}, prints the result on stdout. Inside an execute_code script the model can therefore drive arbitrary tool fan-out without re-entering the LLM: for f in $(pasclaw __tool fs_list '{"path":"src/"}' | jq -r .files[]); do pasclaw __tool memory_search "{\"q\":\"$f\"}"; done. Same registry the model is using, same sandbox + denylist + output-cap path; bad tokens get rejected at the wire level, missing info file produces an actionable error. The killer use case is "list every *.pas under src/, grep each for symbol X, build a CSV" — pre-PR: 1 + N + 1 inference rounds; post-PR: one execute_code call.

Stats overlay (TUI) — type /stats inside the TUI for a read-only modal showing per-session token totals (in/out/cache), turn count + wall-clock elapsed, truncation count + bytes saved, OutputCache handles held, and per-tool call counts (top 8 rendered). Pairs with tool_output_cap so the savings are visible — type /stats after a long session, see the bytes saved climb. Any key dismisses the overlay. The overlay uses an in-process accumulator and works regardless of stats_collection_enabled.

Stats persistence + web UI (gateway) — flip stats_collection_enabled in config.json (or answer "y" during pasclaw onboard) to persist per-turn counters into each session JSON's meta.stats block: input_tokens, output_tokens, cache_read_tokens, cache_created_tokens, turns, tool_calls, truncation_bytes_saved. The gateway exposes a 5-second-cached aggregate at GET /v1/stats returning the same fields summed across every session in $PASCLAW_HOME/workspace/sessions/, plus by_provider and by_model rollups (so the operator can see "OpenAI ate 70% of my tokens this week"). The web UI's Stats tab auto-refreshes every 10s and renders the cards + rollup tables; when the flag is off the panel shows a hint pointing at the config flag rather than a wall of zeros. Default off: per-session JSONs of flag-off operators are byte-identical to the pre-feature schema.

The gateway's own stateless API endpoints (/v1/chat, /v1/chat/completions, /v1/responses) accumulate into per-endpoint synthetic session files (_gateway_v1_chat.json, _gateway_v1_chat_completions.json, _gateway_v1_responses.json) so calls driven through the web UI chat or any other OpenAI/Anthropic-compatible client also show up in /v1/stats. Each endpoint gets one bucket regardless of model — Meta.Model tracks the most recent request's model, so the by_model rollup reflects the latest call rather than per-call breakdown for those buckets (trade-off chosen for simplicity; operators wanting per-model breakdown of API traffic can route through real sessions via pasclaw agent --session <id>). Concurrent gateway requests serialise the bucket update through a single critical section so two simultaneous turns don't race the file.

Auto-router (cheap-tier routing) — flip auto_router.enabled plus name a configured fallback in auto_router.easy_provider (or answer "y" to both onboarding prompts) and PasClaw classifies each user message before the loop fires. Three signals: (1) hard keyword markers like implement, refactor, debug, fix bug, write tests, optimize route hard outright; (2) token-count threshold (auto_router.easy_max_tokens, default 500) — anything longer routes hard; (3) tool-mix gate — when shell_exec / execute_code / fs_write are in the registry, ambiguous one-word messages (yes, continue) abstain because they might mean "now go run the script". Only when all three agree on easy AND the operator has explicitly enabled routing does the agent swap providers for that turn. The original primary is automatically prepended to the fallback chain on a routed turn, so a misclassified easy that the cheap provider fumbles falls back to the primary with no operator intervention. Each routed turn surfaces (routed -> <provider>, auto-routed) in the assistant header so it's never invisible.

Provider authentication and model selection

pasclaw auth status
pasclaw auth login anthropic
pasclaw auth logout anthropic
pasclaw auth weixin

pasclaw model show
pasclaw model set claude-opus-4-7
pasclaw model add openai gpt-4o-mini
pasclaw model refresh anthropic              # GET /v1/models → cache under $PASCLAW_HOME/cache/models/anthropic.json
pasclaw model refresh --all                  # refresh every configured provider in one shot
pasclaw model list openai                    # show the cached roster, newest first

auth login prompts for an API key and stores it in the matching provider entry. model set changes default_model; model add upserts a provider entry and records a model for that provider. model refresh hits the provider's live /v1/models endpoint (Anthropic / OpenAI / Gemini / xAI / Groq / OpenRouter / DeepSeek / Mistral / Moonshot / Cerebras / Ollama / vLLM / LM Studio — anything in the catalog), caches the result, and shows the count. pasclaw onboard runs the same fetch automatically once you've entered the API key and presents a numbered picker over the top 12 newest models — type a number, type a free-form name, or hit Enter for the catalog default. Discovery fails gracefully when the endpoint is unreachable (offline, behind a proxy, provider 5xx) and the onboarding flow falls back to the existing text input.

MCP servers

pasclaw mcp list
pasclaw mcp add filesystem npx -y @modelcontextprotocol/server-filesystem /tmp
pasclaw mcp add remote https://example.com/mcp
pasclaw mcp show filesystem
pasclaw mcp test filesystem
pasclaw mcp remove filesystem
pasclaw mcp edit

pasclaw mcp catalog          # list public MCP servers PasClaw knows about
pasclaw mcp install <name>   # add one from the catalog; auth from $ENV_VAR

MCP entries are stored in the config as mcp_servers. A command starting with http:// or https:// is tested with the HTTP MCP client; other commands are spawned with the stdio MCP client.

Catalog (curated public MCP servers; opt-in, never preloaded). pasclaw mcp install <name> writes a normal mcp_servers entry — list/remove/test/show all just work afterwards. Auth is read from the env var the catalog entry names; installing with the env unset writes an empty Authorization header and a hint to re-run after setting it.

name	env var	provider
replicate	REPLICATE_API_TOKEN	Run AI models (text/image/video/audio) on Replicate — 5000+ models.
digitalocean-apps	DIGITALOCEAN_TOKEN	Manage DigitalOcean App Platform.
digitalocean-databases	DIGITALOCEAN_TOKEN	Manage DigitalOcean Managed Databases.
runpod-docs	(none)	Search RunPod documentation.
huggingface	HF_TOKEN	Search models / datasets / papers / Spaces.

Cron

pasclaw cron list
pasclaw cron add daily-summary "0 9 * * *" summarize "workspace/memory"
pasclaw cron add ping-discord "*/15 * * * *" healthcheck "--channel discord:https://discord.com/api/webhooks/..."
pasclaw cron add line-status "0 * * * *" status_skill "--channel line:U1234abcd"
pasclaw cron disable daily-summary
pasclaw cron enable daily-summary
pasclaw cron remove daily-summary

Each cron entry persists its last successful fire time to $PASCLAW_HOME/workspace/cron/state.json so a missed slot (gateway down, laptop closed) catches up on the next tick instead of being silently skipped. Delivery is at-least-once: the state file is written after the skill runs (PasClaw.Cron.Scheduler.pas:175 runs the skill, :196 persists the timestamp), so a crash between those two steps will replay the job on restart. Idempotent skills are safe; side-effecting skills (sending emails, posting to channels) should self-deduplicate. Skill output is appended to workspace/memory/<today>.md for the model to recall on subsequent turns, and — if --channel <kind>:<target> was set — posted to the configured channel. Channel kinds: discord, slack, teams, webhook (URL is the target), line (target is userId, token from $PASCLAW_LINE_TOKEN), whatsapp (target is phone number, credentials from $PASCLAW_WHATSAPP_TOKEN + $PASCLAW_WHATSAPP_PHONE_ID).

Sessions

pasclaw agent --session 20260601T093015-1a2b3c4d   # resume by id (creates fresh if absent)
pasclaw resume 20260601T093015-1a2b3c4d            # shorthand for the above

pasclaw session list                # id, age, title, msg count for every saved session
pasclaw session show <id>           # metadata + last 8 messages (head trimmed)
pasclaw session export <id>         # raw JSON to stdout (pipe through `jq`)
pasclaw session delete <id>         # remove the session file

Sessions live as one JSON file per id under $PASCLAW_HOME/workspace/sessions/<id>.json. The file carries the full messages array (role / content / name / tool_call_id / tool_calls), the model + provider + title in meta, and any system_prompt_override left behind by /compact. pasclaw agent saves after each turn (atomic .tmp → rename), so killing the process mid-conversation is safe — pasclaw resume <id> picks up exactly where it stopped. /new inside an interactive session spawns a new id; /reset clears history in place and keeps the same id; /compact is persisted before returning to the prompt so a summariser pass survives an immediate /quit. The Delphi build of pasclaw tui uses the same store (session list pane reads ListSessions, the chat pane operates against the loaded TSession.Messages); the FPC build of pasclaw tui stays in-memory.

Web search

Two tools, both registered automatically alongside the filesystem and shell tools:

• web_search(query, k?) — returns up to k results as title + URL + snippet. Dispatches to the configured provider; defaults to DuckDuckGo when nothing is set.
• web_fetch(url, max_chars?) — fetches an http:// or https:// URL and returns the response body as readable plain text (HTML tags stripped, entities decoded, whitespace collapsed, capped at 50 KB by default).

Provider is set under web_search in ~/.pasclaw/config.json:

{
  "web_search": {
    "provider":    "brave",
    "api_key":     "",
    "max_results": 5
  }
}

Provider	API key needed?	Source
duckduckgo (default)	no	HTML scrape of html.duckduckgo.com/html/
brave	yes — $PASCLAW_BRAVE_API_KEY overrides api_key	api.search.brave.com/res/v1/web/search
tavily	yes — $PASCLAW_TAVILY_API_KEY overrides api_key	api.tavily.com/search
searxng	no (most public instances); optional $PASCLAW_SEARXNG_API_KEY for protected ones	<web_search.base_url>/search?format=json
perplexity	yes — $PASCLAW_PERPLEXITY_API_KEY overrides api_key	api.perplexity.ai/chat/completions (Sonar model — returns one synthesised answer plus citation URLs)
gemini	yes — $PASCLAW_GEMINI_API_KEY (or $PASCLAW_GOOGLE_API_KEY) overrides api_key	generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent with google_search grounding — returns one synthesised answer plus the ground-truth source URLs Gemini consulted

Env-var values win over the api_key field so secrets can stay out of config.json. SearXNG additionally needs web_search.base_url set since every instance is self-hosted (e.g. "base_url": "https://searx.be").

Skills

A skill is a markdown manifest under $PASCLAW_HOME/workspace/skills/. The system prompt lists each installed skill (name + one-line description + path); the model reads the full body with fs_read when a matching task comes up. Skills tagged kind: shell or kind: prompt also register as a callable skill_<name> tool.

pasclaw skills list
pasclaw skills install owner/repo                         # GitHub repo root
pasclaw skills install owner/repo/path/to/skill           # GitHub subdirectory
pasclaw skills install owner/repo/path/to/[email protected]    # GitHub at a pinned ref
pasclaw skills install clawhub:code-review                # ClawHub: latest version
pasclaw skills install clawhub:[email protected]          # ClawHub: pinned version
pasclaw skills search "code review"                       # ClawHub: search the registry
pasclaw skills install my-skill                           # Legacy: record name in config.json
pasclaw skills remove my-skill                            # Delete workspace dir + config entry

On-disk layout

PasClaw accepts two layouts:

• Per-directory SKILL.md (preferred — same format picoclaw, nanobot, ClawHub, and Anthropic agent-skills use):

  workspace/skills/my-skill/
  └── SKILL.md     ← YAML frontmatter + markdown body
  

  ---
  name: my-skill
  description: One-line summary the model uses to pick the skill
  # Omit `kind` for knowledge-only skills (most common); set `kind: shell`
  # or `kind: prompt` to register a callable `skill_<name>` tool.
  ---
  
  # My skill
  
  Markdown body. The system prompt advertises this SKILL.md path; the
  model loads the full body with `fs_read` when the matching task comes
  up.
  

  A copy-pasteable starter lives at samples/skills/hello/SKILL.md.

• Legacy single *.json (workspace/skills/<name>.json) — still loaded for backwards compat. Per-directory SKILL.md entries shadow same-named JSON entries; new skills should use the directory layout. The JSON shape mirrors the frontmatter:

  {
    "name": "my-skill",
    "description": "One-line summary the model uses to pick the skill",
    "kind": "shell",
    "schema": "{\"type\":\"object\",\"properties\":{\"path\":{\"type\":\"string\"}}}"
  }
  

  kind and schema are only needed for callable skills; knowledge-only skills carry just name + description.

Install from GitHub

pasclaw skills install owner/repo[/path][@ref]:

• Fetches a zip snapshot from codeload.github.com.
• Extracts it via the bundled zip library — Zipper.TUnZipper under FPC, System.Zip.TZipFile under Delphi. No tar dependency.
• Locates SKILL.md at the requested subpath and validates it through ParseSkillMD.
• Copies the containing directory tree into $PASCLAW_HOME/workspace/skills/<dest>/, where <dest> is the last segment of the subpath, or the repo name when no subpath was given.
• When @ref is omitted, tries main first and falls back to master.
• Refuses to overwrite an existing skill directory — run pasclaw skills remove <name> first to reinstall.

Install from ClawHub

pasclaw skills install clawhub:<slug>[@<version>] talks to ClawHub (https://clawhub.ai), the slug-based registry picoclaw and nanobot standardised on:

• GET /api/v1/skills/<slug> — fetches metadata, surfaces moderation flags, and resolves latestVersion when no @<version> is pinned.
• GET /api/v1/download?slug=<slug>&version=<version> — pulls the zip, then runs it through the same PasClaw.Skills.Zip + ParseSkillMD validation pipeline as the GitHub install path.
• Malware-flagged skills are refused; skills flagged as suspicious install with a warning.
• Slugs are lowercase alphanumerics with - or _.
• The clawhub: prefix is required. A bare slug-shaped name like my-skill still falls through to the legacy config.json-only record, so pre-Phase-3 install scripts keep working unchanged.

pasclaw skills search <query> hits /api/v1/search and prints slug / version / display-name / summary rows.

Subsequent phases will add scripts/ (callable helpers) + references/ (lazy-loaded context) runtime support.

Gateway, OpenAI-compatible server, and channels

pasclaw gateway
pasclaw gateway --addr 0.0.0.0 --port 8088
pasclaw gateway --telegram --token <BOT_TOKEN>
pasclaw gateway --line                              # also pass $PASCLAW_LINE_TOKEN + $PASCLAW_LINE_SECRET
pasclaw gateway --whatsapp                          # also pass $PASCLAW_WHATSAPP_{TOKEN,PHONE_ID,VERIFY_TOKEN,APP_SECRET}
pasclaw gateway --matrix                            # also pass $PASCLAW_MATRIX_HOMESERVER + $PASCLAW_MATRIX_TOKEN
pasclaw gateway --irc                               # also pass $PASCLAW_IRC_{SERVER,NICK,CHANNEL}
pasclaw gateway --no-tools --no-mcp --no-hashline

pasclaw serve
pasclaw serve --addr 0.0.0.0 --port 8088
pasclaw serve --debug
pasclaw serve --max-iter 40
pasclaw serve --no-tools --no-mcp --no-hashline

Both commands use the same TGatewayServer implementation. gateway starts the full surface and can also run the Telegram long-poll channel. serve is a focused wrapper for OpenAI-compatible clients and prints copy-pasteable client configuration.

OpenAI-compatible client example:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8088/v1", api_key="sk-pasclaw")
response = client.chat.completions.create(
    model="claude-opus-4-7",
    messages=[{"role": "user", "content": "hello"}],
)
print(response.choices[0].message.content)

Curl examples:

curl http://127.0.0.1:8088/v1/health
curl http://127.0.0.1:8088/v1/tools
curl http://127.0.0.1:8088/v1/chat \
  -H 'Content-Type: application/json' \
  -d '{"message":"hello"}'
curl http://127.0.0.1:8088/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7","messages":[{"role":"user","content":"hello"}]}'
curl http://127.0.0.1:8088/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7","input":"hello"}'

The gateway routes implemented in src/pkg/gateway/ are:

Route	Method	Purpose
/	GET	Embedded single-page web UI from src/pkg/gateway/webui.html. Tabs for chat (with streaming + tool-call rendering + localStorage session history), memory browser, file browser, MCP servers, cron entries, skills, log tail (SSE), and a read-only config viewer.
/v1	GET	JSON index listing gateway routes.
/v1/health	GET	Health check with PasClaw version.
/v1/version	GET	Version and build metadata.
/v1/status	GET	Default provider/model plus provider, MCP, cron, skill, and tool counts.
/v1/tools	GET	Registered tool descriptors.
/v1/mcp	GET	Configured MCP servers (name, cmd, args, enabled).
/v1/cron	GET	Cron entries (id, spec, skill, args, channel_*, enabled).
/v1/skills	GET	Installed skills (name, description, kind, path, dir).
/v1/memory	GET	Files in workspace/memory/ with sizes.
/v1/memory/<name>	GET	Contents of one memory file (rejects path-traversal).
/v1/config	GET	Full config with providers[].api_key masked to •••.
/v1/fs?path=…	GET	Directory listing (entries + sizes + dir-flag); defaults to $PASCLAW_HOME.
/v1/fs/read?path=…	GET	File contents capped at 256 KB; truncated flag on response.
/v1/logs	GET	SSE tail of the gateway log buffer (1000-entry ring); recent buffer dumps first, then live.
/v1/chat	POST	PasClaw JSON chat endpoint accepting {"message":"..."}.
/v1/chat/completions	POST	OpenAI Chat Completions-compatible endpoint; supports streaming with stream: true.
/v1/responses	POST	OpenAI Responses-compatible endpoint accepting string or message-array input; non-streaming only.
/v1/models	GET	OpenAI-compatible model list containing the configured default model.

When /v1/chat/completions runs with stream: true, the tool loop executes
server-side and each tool call is surfaced to the client as a visible
content delta in a Claude-Code-style transcript — the tool name with its key
argument, followed by a short result summary on the next line:

⏺ fs_read(README.md)
  ⎿ 312 lines, 12044 bytes — ¶README.md#a1b2
⏺ shell_exec(ls -la)
  ⎿ exit=0

Known tools (fs_read, fs_write, fs_list, fs_grep,
fs_edit_hashline, shell_exec, memory_search, web_search, web_fetch)
surface their most meaningful argument;
MCP and other tools fall back to a compact one-line dump of the raw
arguments. The full argument and result text also go to the SSE comment
lines (: tool_call ... / : tool_result ...) for consumers that log
structured activity, and to the server debug log when --debug is set.
The formatter lives in src/pkg/gateway/PasClaw.Gateway.ToolView.pas
(unit-tested via make test-toolview).

Manual /v1/responses verification examples:

# string input
curl http://127.0.0.1:8088/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7","input":"hello"}'

# message-array input
curl http://127.0.0.1:8088/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7","input":[{"role":"user","content":[{"type":"input_text","text":"hello"}]}]}'

# missing or empty input should return invalid_request_error
curl http://127.0.0.1:8088/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7","input":""}'

# streaming is intentionally unsupported on this route
curl http://127.0.0.1:8088/v1/responses \
  -H 'Content-Type: application/json' \
  -d '{"model":"claude-opus-4-7","input":"hello","stream":true}'

Channel posting commands:

pasclaw post discord <webhook-url> "hello"
pasclaw post slack <webhook-url> "hello"

Embedding in your own app

TPasClawAgent (unit PasClaw.Agent) is a TComponent you can drop on a form or instantiate from code to embed the full agent — provider auth, tools, MCP, skills — inside a standalone Delphi or FPC binary, without shelling out to the CLI.

Code-driven (one-liner) form:

uses PasClaw.Agent, PasClaw.Tools;

var
  Agent: TPasClawAgent;
begin
  Agent := TPasClawAgent.Create('claude-opus-4-7');
  try
    Agent.SetProvider('anthropic', GetEnvironmentVariable('ANTHROPIC_API_KEY'));
    Agent.RegisterTool(TWebSearchTool.Create);
    Agent.RegisterTool(TWebFetchTool.Create);
    Agent.RegisterTool(TFileSystemTool.Create);
    WriteLn(Agent.Run('Summarize the latest Delphi release notes.'));
  finally
    Agent.Free;
  end;
end;

SetProvider(Kind, APIKey) builds a TProviderConfig in-memory from the catalog entry for Kind (anthropic, openai, gemini, groq, ollama, etc. — see PasClaw.Providers.Catalog) plus the API key the caller hands in. The catalog supplies the default base URL and default model; both can be overridden via the three-arg and four-arg SetProvider overloads. This lets an embedded binary run without ever touching ~/.pasclaw/config.json — set the env var, ship the binary, done.

Run raises EPasClawRun on failure; use Chat(prompt, reply, err): Boolean when you'd rather not unwind exceptions. RegisterTool takes ownership of the TPasClawTool instance and frees it with the agent.

Form-designer / property-driven form (back-compat — Create(nil), set published properties, wire events):

PC := TPasClawAgent.Create(nil);
PC.Model       := 'claude-opus-4-7';
PC.SystemPrompt := 'You are a Pascal expert.';
PC.OnText := MyTextHandler;
if PC.Chat('hi', Reply, Err) then WriteLn(Reply);

Built-in tools available as classes: TWebSearchTool, TWebFetchTool, TFileSystemTool (bundle: fs_read/write/grep/list/edit), TShellTool, TMemoryTool. Custom tools subclass TPasClawTool (unit PasClaw.Tools.Obj) and override Name / Description / Schema / Run.

TPasClawServer (same unit) hosts the full HTTP gateway inside the calling process. Same OOP shape as TPasClawAgent:

uses PasClaw.Agent, PasClaw.Tools;

var
  Server: TPasClawServer;
begin
  Server := TPasClawServer.Create('0.0.0.0', 8088);
  try
    Server.SetProvider('anthropic', GetEnvironmentVariable('ANTHROPIC_API_KEY'));
    Server.RegisterTool(TWebSearchTool.Create);
    Server.Run;  { blocks until Stop is signalled from another thread }
  finally
    Server.Free;
  end;
end;

SetProvider works the same way as on TPasClawAgent — call it before Start and the gateway boots with the in-memory provider config.

Run does Start + WaitForStop in one call and raises EPasClawRun if startup fails. Use Start / WaitForStop / Stop separately when you need to do something between binding the socket and entering the wait. SIGINT handling is the embedder's problem — most hosting apps already have their own signal-handling strategy, so the component doesn't install one.

samples/component-console/ ships three sample binaries:

cd samples/component-console
make get-indy   # only needed once, from repo root
make            # builds all three (SampleConsole, SampleSimple, SampleServer)
make simple     # just the agent code-form sample
make server     # just the server sample

PasClaw.Component is still available as the legacy unit name — it now re-exports everything from PasClaw.Agent, so existing code keeps compiling.

Knowledgebase (RAG over your documents)

Index reference documents — manuals, books, source trees — so the agent can
retrieve them while answering ("Big RAG" style, but operator-curated):

pasclaw kb add ~/docs/delphi-book.md ~/projects/mylib/   # files and/or dirs
pasclaw kb list
pasclaw kb search "constructor constraints in generics"
pasclaw kb sync          # after documents change
pasclaw kb status
pasclaw kb remove ~/projects/mylib/

Documents are indexed in place (never copied) into workspace/kb.db,
chunked by paragraph (~1.6 KB target). Supported types: markdown, plain
text, HTML (tag-stripped), and common source-code extensions (.pas,
.dpr, .c, .py, ...). PDFs are not supported — convert first, e.g.
pdftotext book.pdf book.txt.

Once at least one source is registered, two read-only agent tools appear
automatically (and not before, so the model never sees an empty KB):

Tool	Purpose
kb_search(query, k)	Search the corpus; returns path#cN citations + snippets + scores.
kb_get(path, chunk, window)	Expand a citation into full chunk text ± window neighbouring chunks.

Ranking is SQLite FTS5 BM25 out of the box. When the local embedding
runtime is provisioned (pasclaw memory provision — the same sqlite-vec +
ONNX Runtime + MiniLM artifacts memory_search uses, gated by the same
vector_search_enabled flag), the KB automatically upgrades to hybrid
keyword + semantic ranking with Reciprocal Rank Fusion; embeddings are
computed locally and never leave the host. The vector sidecar
(workspace/kb.db.vec) is rebuilt by kb sync when documents change.
Unlike memory_search, the KB does not re-index on every query — kb sync is the explicit refresh, keeping kb_search fast on large corpora.

Conversation memory (memory_search) and the knowledgebase are deliberately
separate: one answers "what did we decide earlier", the other "what do my
reference documents say".

Maintenance and diagnostics

pasclaw status
pasclaw version
pasclaw migrate
pasclaw update --check
pasclaw update --repo owner/name
pasclaw membench --records 1000 --content 128
pasclaw membench --records 1000 --content 128 --keep --out /tmp

Providers

Provider definitions live in src/pkg/providers/PasClaw.Providers.Catalog.pas. pasclaw onboard uses the catalog to populate provider kind, default API base URL, default model, and auth scheme.

Implemented protocol families:

• pfAnthropic: Anthropic Messages API using x-api-key.
• pfOpenAI: OpenAI Chat Completions-compatible HTTP API. This powers OpenAI and most compatible hosted/local providers.
• pfGemini: Google Gemini generateContent REST API using x-goog-api-key.
• pfPlaceholder: catalog placeholder for known future provider families that are not wired into this build.

Auth schemes:

• asBearer: Authorization: Bearer <key>.
• asNone: no auth header, used for local Ollama/vLLM deployments.
• asHeader: raw key in the catalog-specified header name.

Current catalog entries:

Provider	kind	Family	Default base	Default model	Auth
Anthropic	anthropic	Anthropic	https://api.anthropic.com	claude-opus-4-7	x-api-key header
OpenAI	openai	OpenAI-compatible	https://api.openai.com	gpt-4o-mini	Bearer
OpenRouter	openrouter	OpenAI-compatible	https://openrouter.ai/api	provider-selected	Bearer
Zhipu (GLM)	zhipu	OpenAI-compatible	https://open.bigmodel.cn/api/paas	glm-4	Bearer
DeepSeek	deepseek	OpenAI-compatible	https://api.deepseek.com	deepseek-chat	Bearer
Volcengine (Doubao/Ark)	volcengine	OpenAI-compatible	https://ark.cn-beijing.volces.com/api	provider-selected	Bearer
Qwen (DashScope)	qwen	OpenAI-compatible	https://dashscope.aliyuncs.com/compatible-mode	qwen-max	Bearer
Groq	groq	OpenAI-compatible	https://api.groq.com/openai	provider-selected	Bearer
Moonshot (Kimi)	moonshot	OpenAI-compatible	https://api.moonshot.cn	moonshot-v1-32k	Bearer
MiniMax	minimax	OpenAI-compatible	https://api.minimax.chat	provider-selected	Bearer
Mistral	mistral	OpenAI-compatible	https://api.mistral.ai	mistral-large-latest	Bearer
NVIDIA NIM	nvidia	OpenAI-compatible	https://integrate.api.nvidia.com	provider-selected	Bearer
Cerebras	cerebras	OpenAI-compatible	https://api.cerebras.ai	provider-selected	Bearer
Novita AI	novita	OpenAI-compatible	https://api.novita.ai	provider-selected	Bearer
Xiaomi MiMo	mimo	OpenAI-compatible	set api_base in config	provider-selected	Bearer
Ollama (local)	ollama	OpenAI-compatible	http://localhost:11434	required during onboarding	none
vLLM (local)	vllm	OpenAI-compatible	http://localhost:8000	required during onboarding	none
LiteLLM proxy	litellm	OpenAI-compatible	set api_base in config	provider-selected	Bearer
Google Gemini	gemini	Gemini	https://generativelanguage.googleapis.com	gemini-1.5-flash	x-goog-api-key header

Minimal local provider examples:

{
  "default_provider": "ollama",
  "default_model": "llama3.1:8b",
  "gateway": { "log_level": "info", "bind_addr": "127.0.0.1", "port": 8088 },
  "providers": [
    {
      "name": "ollama",
      "kind": "ollama",
      "api_base": "http://localhost:11434",
      "api_key": "",
      "model": "llama3.1:8b"
    }
  ],
  "mcp_servers": [],
  "crons": [],
  "skills": []
}

{
  "default_provider": "vllm",
  "default_model": "your-model",
  "gateway": { "log_level": "info", "bind_addr": "127.0.0.1", "port": 8088 },
  "providers": [
    {
      "name": "vllm",
      "kind": "vllm",
      "api_base": "http://localhost:8000",
      "api_key": "",
      "model": "your-model"
    }
  ],
  "mcp_servers": [],
  "crons": [],
  "skills": []
}

🔒 Security / sandbox

PasClaw's filesystem and shell tools are guarded by an opt-in workspace boundary plus an always-on shell denylist (src/pkg/tools/PasClaw.Tools.Sandbox.pas). Configure via the sandbox block in ~/.pasclaw/config.json:

"sandbox": {
  "restrict_to_workspace":        true,
  "allow_read_outside_workspace": false,
  "workspace":                    "/home/me/my-project",
  "allow_read_paths":  ["^/usr/(include|share)/.*"],
  "allow_write_paths": ["^/tmp/agent/.*"],
  "custom_shell_deny": ["scp ", "rsync "],
  "shell_deny_enabled":           true
}

Field	Default	Effect
restrict_to_workspace	false	When true, fs_read / fs_write / fs_list / fs_edit_hashline / fs_grep refuse paths outside workspace. shell_exec refuses absolute paths outside it AND tokens containing .., pins the shell's cwd to the workspace, and bans cd / chdir / pushd / popd.
allow_read_outside_workspace	false	When true, reads are allowed anywhere even while writes stay restricted. Useful for letting the agent pull from /usr/include/ while still locking down writes.
workspace	"" (cwd at startup)	Absolute path the agent may operate inside. Empty means "use the current working directory at the time pasclaw was invoked", which is the picoclaw / Claude Code convention.
allow_read_paths	[]	PCRE regex patterns that also count as readable. Same syntax picoclaw's tools.allow_read_paths accepts — anchors (^ $), character classes, alternation.
allow_write_paths	[]	Same for writes.
custom_shell_deny	[]	Extra substrings appended to the built-in shell denylist. Case-insensitive.
shell_deny_enabled	true	Master switch for the shell denylist. Set false only for trusted automation — doing so re-enables sudo, rm, dd, mkfs, $( ), curl | sh, format c:, PowerShell -EncodedCommand, etc.
block_private_networks	true	When true, web_fetch refuses URLs whose host resolves to a private / loopback / link-local IPv4 address (RFC1918, 127.0.0.0/8, 169.254.0.0/16 — including the cloud-metadata endpoint 169.254.169.254, CGNAT, IETF-reserved). Initial URL and every redirect hop are both checked. Flip to false only when you actually need the model to reach private addresses. See PasClaw.Net.SSRF for the full blocklist.

Cross-target regex: PasClaw.Tools.Regex wraps FPC's RegExpr and Delphi's System.RegularExpressions behind one call, so allow_*_paths patterns are full PCRE on either toolchain. Invalid patterns return False (the sandbox falls through to the workspace boundary) rather than crashing.

Built-in shell denylist (always on unless shell_deny_enabled is false):

• POSIX tokens: sudo, su, rm, chmod, chown, pkill, killall, kill, shutdown, reboot, poweroff, halt, eval, mkfs, diskpart.
• Windows tokens: del, erase, rd, rmdir, format, attrib, takeown, icacls, runas.
• cwd-change tokens (when restrict_to_workspace): cd, chdir, pushd, popd. Any token containing .. is also rejected.
• Substrings: dd if=, :(){:|, <<EOF, $( ), ${ }, backticks, | sh, | bash, apt install/remove/purge, yum install/remove, dnf install/remove, npm install -g, pip install --user, docker run/exec, git push, git force, format c:.
• PowerShell (matched lowercased): powershell -e/-en/-enc/-ec, -encodedcommand, iex (, invoke-expression, [convert]::frombase64, [text.encoding], .getstring([byte[], set-executionpolicy.
• Device writes: > /dev/sd* / /hd* / /vd* / /xvd* / /nvme* / /mmcblk* / /loop* / /md*.
• Always-safe paths: /dev/null, /dev/zero, /dev/{,u}random, /dev/std{in,out,err} — picoclaw's safePaths.

Workspace-pin: when restrict_to_workspace=true, Tool_Shell invokes RunOneShot with WorkingDir = workspace so the child shell starts inside the boundary. Combined with the cd token ban and .. traversal check, a sandboxed model has no relative-path escape — even if a future denylist gap let a command through, the shell still starts in the workspace, not wherever pasclaw was launched from.

Known limitation: path canonicalisation uses ExpandFileName, which resolves .. but not symlinks. Picoclaw's equivalent (os.OpenRoot in Go 1.24+) enforces the boundary at the syscall layer; PasClaw runs on FPC and Delphi RTLs that have no equivalent. Do not place symlinks inside workspace that point outside it — they would let the agent escape.

--no-tools remains the strongest option: it disables the tool registry entirely, so neither fs_* nor shell_exec is registered. The system prompt automatically reflects this (no SKILLS section, no "ALWAYS use tools" rule).

Repository layout

src/
  pasclaw/          Program entry point (`PasClaw.dpr`)
  cmd/              CLI command units and root dispatcher
  pkg/
    agent/          Agent execution and prompts
    channels/       Telegram, Discord, Slack, Teams, generic webhook, LINE, WhatsApp, Matrix, IRC
    cliui/          ANSI styling, banner, command help rendering
    component/      Shared components
    config/         Version constants and on-disk config model
    cron/           Cron scheduler
    gateway/        Indy HTTP server, OpenAI-compatible API, embedded web UI
    hashline/       Hashline patch/edit support
    json/           Project JSON abstraction
    logger/         Levelled logging
    mcp/            MCP stdio/HTTP clients and tool bridge
    membench/       Memory benchmark helpers
    memory/         Memory log storage
    net/            SSRF guard (IPv4 blocklist + DNS re-resolution)
    platform/       Platform helpers
    providers/      Provider catalog and LLM HTTP clients
    search/         Web-search providers (DuckDuckGo, Brave, Tavily, SearXNG, Perplexity, Gemini) + HTML→text
    skills/         Skill manifest loading and tool registration
    tokenizer/      Token counting helpers
    tools/          Built-in tool registry, filesystem, shell, and tool loop
    tui/            Full-screen terminal UI
    updater/        GitHub release update support
    utils/          Path, file, and string helpers

License

MIT. See LICENSE.