ductor

Claude Code, Codex CLI, and Gemini CLI as your coding assistant — on Telegram and Matrix.
Uses only official CLIs. Nothing spoofed, nothing proxied. Multi-transport, automation, and sub-agents in one runtime.

Quick start · How chats work · Commands · Docs · Contributing

If you want to control Claude Code, Google's Gemini CLI, or OpenAI's Codex CLI via Telegram or Matrix, build automations, or manage multiple agents easily — ductor is the right tool for you. The messaging layer is modular: Telegram and Matrix ship today, and new transports plug into the same transport-agnostic core.

ductor runs on your machine and sends simple console commands as if you were typing them yourself, so you can use your active subscriptions (Claude Max, etc.) directly. No API proxying, no SDK patching, no spoofed headers. Just the official CLIs, executed as subprocesses, with all state kept in plain JSON and Markdown under ~/.ductor/.

Quick start

pipx install ductor    # or: uv tool install ductor
ductor

The onboarding wizard handles CLI checks, transport setup (Telegram or Matrix), timezone, optional Docker, and optional background service install.

Requirements: Python 3.11+, at least one CLI installed (claude, codex, or gemini), and either:

• a Telegram Bot Token from @BotFather, or
• a Matrix account on a homeserver (homeserver URL, user ID, password/access token)

For Matrix support: ductor install matrix — see Matrix setup guide.

Detailed setup: docs/installation.md

New in v0.16.0

• Memory maintenance is now built in — streaming compaction boundaries can trigger a silent memory flush, optional reflection hook, and LLM-driven MAINMEMORY.md compaction.
• Telegram UX is tighter — stage-based emoji status reactions are enabled by default, while seen_reaction remains available as the simpler one-shot alternative.
• Lifecycle notifications are routable — startup and upgrade notices can be pinned to specific chats/topics instead of always broadcasting.
• Media transcription is extensible — bundled media tools now accept external audio/video transcription commands via config-driven env-var hand-off.
• Task and multi-agent automation got sharper — background tasks support priorities, and ask_agent_async.py now supports --reply-to and --silent for cleaner pipelines.

Release summary: docs/release_notes_v0.16.0.md

How chats work

ductor gives you multiple ways to interact with your coding agents. Each level builds on the previous one.

1. Single chat (your main agent)

This is where everyone starts. You get a private 1:1 chat with your bot (Telegram or Matrix). Every message goes to the CLI you have active (claude, codex, or gemini), responses stream back in real time.

You:   "Explain the auth flow in this codebase"
Bot:   [streams response from Claude Code]

You:   /model
Bot:   [interactive model/provider picker]

You:   "Now refactor the parser"
Bot:   [streams response, same session context]

This single chat is all you need. Everything else below is optional.

2. Groups with topics (multiple isolated chats)

Telegram: Create a group, enable topics (forum mode), and add your bot.
Matrix: Invite the bot to multiple rooms — each room is its own context.

Every topic (Telegram) or room (Matrix) becomes an isolated chat with its own CLI context.

Group: "My Projects"
  ├── General           ← own context (isolated from your single chat)
  ├── Topic: Auth       ← own context
  ├── Topic: Frontend   ← own context
  ├── Topic: Database   ← own context
  └── Topic: Refactor   ← own context

That's 5 independent conversations from a single group. Your private single chat stays separate too — 6 total contexts, all running in parallel.

Each topic can use a different model. Run /model inside a topic to change just that topic's provider.

All chats share the same ~/.ductor/ workspace — same tools, same memory, same files. The only thing isolated is the conversation context.

Telegram note: The Bot API has no method to list existing forum topics.
ductor learns topic names from forum_topic_created and forum_topic_edited
events — pre-existing topics show as "Topic #N" until renamed.
This is a Telegram limitation, not a ductor limitation.

3. Named sessions (extra contexts within any chat)

Need to work on something unrelated without losing your current context? Start a named session. It runs inside the same chat but has its own CLI conversation.

You:   "Let's work on authentication"        ← main context builds up
Bot:   [responds about auth]

/session Fix the broken CSV export            ← starts session "firmowl"
Bot:   [works on CSV in separate context]

You:   "Back to auth — add rate limiting"     ← main context is still clean
Bot:   [remembers exactly where you left off]

@firmowl Also add error handling              ← follow-up to the session

Sessions work everywhere — in your single chat, in group topics, in sub-agent chats. Think of them as opening a second terminal window next to your current one.

4. Background tasks (async delegation)

Any chat can delegate long-running work to a background task. You keep chatting while the task runs autonomously. When it finishes, the result flows back into your conversation.

You:   "Research the top 5 competitors and write a summary"
Bot:   → delegates to background task, you keep chatting
Bot:   → task finishes, result appears in your chat

You:   "Delegate this: generate reports for all Q4 metrics"
Bot:   → explicitly delegated, runs in background
Bot:   → task has a question? It asks the agent → agent asks you → you answer → task continues

Each task gets its own memory file (TASKMEMORY.md) and can be resumed with follow-ups.

5. Sub-agents (fully isolated second agent)

Sub-agents are completely separate bots — own chat, own workspace, own memory, own CLI auth, own config settings (heartbeat, timeouts, model defaults, etc.). Each sub-agent can use a different transport (e.g. main on Telegram, sub-agent on Matrix).

ductor agents add codex-agent    # creates a new bot (needs its own BotFather token)

Your main chat (Claude):        "Explain the auth flow"
codex-agent chat (Codex):       "Refactor the parser module"

Sub-agents live under ~/.ductor/agents/<name>/ with their own workspace, tools, and memory — fully isolated from the main agent.

You can delegate tasks between agents:

Main chat:  "Ask codex-agent to write tests for the API"
  → Claude sends the task to Codex
  → Codex works in its own workspace
  → Result flows back to your main chat

Comparison

How it all fits together

~/.ductor/                          ← shared workspace (tools, memory, files)
  │
  ├── Single chat                   ← main agent, private 1:1
  │     ├── main context
  │     └── named sessions
  │
  ├── Group: "My Projects"          ← same agent, same workspace
  │     ├── General (own context)
  │     ├── Topic: Auth (own context, own model)
  │     ├── Topic: Frontend (own context)
  │     └── each topic can have named sessions too
  │
  └── agents/codex-agent/           ← sub-agent, fully isolated workspace
        ├── own single chat
        ├── own group support
        ├── own named sessions
        └── own background tasks

Features

• Multi-transport — run Telegram and Matrix simultaneously, or pick one
• Multi-language — UI in English, Deutsch, Nederlands, Français, Русский, Español, Português
• Real-time streaming — live message edits (Telegram) or segment-based output (Matrix)
• Provider switching — /model to change provider/model (never blocks, even during active processes)
• Persistent memory — plain Markdown files that survive across sessions
• Memory maintenance — pre-compaction flush, optional reflection cadence, and LLM-driven compaction
• Cron jobs — in-process scheduler with timezone support, per-job overrides, result routing to originating chat
• Webhooks — wake (inject into active chat) and cron_task (isolated task run) modes
• Heartbeat — proactive checks with per-target settings, group/topic support, chat validation
• Image processing — auto-resize and WebP conversion for incoming images (configurable)
• Media transcription hooks — configurable external audio/video transcription commands for bundled media tools
• Notification routing — startup/upgrade lifecycle messages can target specific chats/topics
• Task priorities — interactive, background, and batch scheduling modes for background work
• Telegram status reactions — stage-aware emoji tracker on the user message while the agent works
• Config hot-reload — most settings update without restart (including language, scene, image)
• Docker sandbox — optional sidecar container with configurable host mounts
• Service manager — Linux (systemd), macOS (launchd), Windows (Task Scheduler)
• Cross-tool skill sync — shared skills across ~/.claude/, ~/.codex/, ~/.gemini/

Messenger support

Telegram is the primary transport — full feature set, battle-tested, zero extra dependencies.

Both transports can run in parallel on the same agent:

{"transport": "telegram"}
{"transport": "matrix"}
{"transports": ["telegram", "matrix"]}

Modular transport architecture

Each messenger is a self-contained module under messenger/<name>/ implementing a
shared BotProtocol. The core (orchestrator, sessions, CLI, cron, etc.) is completely
transport-agnostic — it never knows which messenger delivered the message.

Adding a new messenger (Discord, Slack, Signal, ...) means implementing BotProtocol
in a new sub-package and registering it — the rest of ductor works without changes.
Guide: docs/modules/messenger.md

Auth

Telegram

ductor uses a dual-allowlist model. Every message must pass both checks.

• allowed_user_ids — Telegram user IDs that may talk to the bot. At least one required.
• allowed_group_ids — Telegram group IDs where the bot may operate. Default [] = no groups.
• group_mention_only — When true, the bot only responds in groups when @mentioned or replied to.

All three are hot-reloadable — edit config.json and changes take effect within seconds.

Privacy Mode: Telegram bots have Privacy Mode enabled by default and only see /commands in groups. To let the bot see all messages, make it a group admin or disable Privacy Mode via BotFather (/setprivacy → Disable). If changed after joining, remove and re-add the bot.

Group management: When the bot is added to a group not in allowed_group_ids, it warns and auto-leaves. Use /where to see tracked groups and their IDs.

Channel allowlist: Telegram channels are tracked separately via allowed_channel_ids. Unauthorized channels are announced and auto-left on join/audit just like unauthorized groups.

Tip — adding a group for the first time:

1. Create a Telegram group, enable topics if you want isolated chats
2. Add the bot and make it admin (required for full message access)
3. Send a message mentioning @your_bot — the bot won't respond yet
4. In your private chat with the bot, run /where — you'll see the group listed under "Rejected" with its ID
5. Tell the bot: "Add this as an allowed group in the config" — it updates config.json for you
6. Run /restart — the bot now responds in the group

Matrix

Matrix auth uses room and user allowlists in the matrix config block:

• allowed_rooms — Room IDs or aliases where the bot may operate.
• allowed_users — Matrix user IDs allowed to interact with the bot.

group_mention_only nuance on Matrix:

• In non-DM rooms, when group_mention_only=true, the bot requires @mention/reply and bypasses allowed_users checks for those group messages.
• Room-level filtering (allowed_rooms) still applies.

The bot logs in with password on first start, then persists access_token and device_id for subsequent runs. E2EE is supported via matrix-nio[e2e].

Language

ductor's UI (commands, status messages, onboarding) is available in multiple languages:

Set the language in config.json:

{"language": "de"}

This is hot-reloadable — change the language without restarting the bot.

Commands

/new is intentionally a factory reset for the current SessionKey: it clears the bucket tied to the configured default model/provider for that chat or topic, not whichever provider you last switched to temporarily via /model.

Common CLI commands

ductor                  # Start bot (auto-onboarding if needed)
ductor onboarding       # Re-run setup wizard
ductor reset            # Full reset + onboarding
ductor stop             # Stop bot
ductor restart          # Restart bot
ductor upgrade          # Upgrade and restart
ductor status           # Runtime status
ductor help             # CLI overview
ductor uninstall        # Remove bot + workspace

ductor service install  # Install as background service
ductor service status   # Show service status
ductor service start    # Start service
ductor service stop     # Stop service
ductor service logs     # View service logs
ductor service uninstall

ductor docker enable    # Enable Docker sandbox
ductor docker rebuild   # Rebuild sandbox container
ductor docker mount /p  # Add host mount
ductor docker extras    # List optional sandbox packages

ductor agents list      # List configured sub-agents
ductor agents add NAME  # Add a sub-agent
ductor agents remove NAME

ductor api enable       # Enable WebSocket API (beta)
ductor api disable      # Disable WebSocket API

ductor install matrix   # Install Matrix transport extra
ductor install api      # Install API/PyNaCl extra

ductor agents add currently scaffolds Telegram sub-agents interactively. Matrix
sub-agents are supported at runtime, but you configure them via agents.json or
the bundled agent tool scripts.

Workspace layout

~/.ductor/
  config/config.json                 # Bot configuration
  sessions.json                      # Chat session state
  named_sessions.json                # Named background sessions
  tasks.json                         # Background task registry
  cron_jobs.json                     # Scheduled tasks
  webhooks.json                      # Webhook definitions
  agents.json                        # Sub-agent registry (optional)
  SHAREDMEMORY.md                    # Shared knowledge across all agents
  CLAUDE.md / AGENTS.md / GEMINI.md  # Rule files
  logs/agent.log
  workspace/
    memory_system/MAINMEMORY.md      # Persistent memory
    cron_tasks/ skills/ tools/       # Scripts and tools
    tasks/                           # Per-task folders
    telegram_files/ matrix_files/    # Media files (per transport)
    api_files/                       # Uploaded/downloadable API files
    output_to_user/                  # Generated deliverables
  agents/<name>/                     # Sub-agent workspaces (isolated)

Full config reference: docs/config.md — full example with all options: config.example.json

Documentation

Why ductor?

Other projects manipulate SDKs or patch CLIs and risk violating provider terms of service. ductor simply runs the official CLI binaries as subprocesses — nothing more.

• Official CLIs only (claude, codex, gemini)
• Rule files are plain Markdown (CLAUDE.md, AGENTS.md, GEMINI.md)
• Memory is one Markdown file per agent
• All state is JSON — no database, no external services

Disclaimer

ductor runs official provider CLIs and does not impersonate provider clients. Validate your own compliance requirements before unattended automation.

• Anthropic Terms
• OpenAI Terms
• Google Terms

Contributing

git clone https://github.com/PleasePrompto/ductor.git
cd ductor
uv sync --extra dev

Run checks with just:

just check   # linters + type checks (parallel)
just test    # test suite
just fix     # auto-fix formatting and lint issues

Or directly with uv:

uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run mypy ductor_bot

Zero warnings, zero errors.

License

MIT