KinetAios

🌐 官网 / Website → https://phinn.github.io/KinetAios/

🚀 We're live on Product Hunt today — upvote & say hi

📊 查看完整四引擎对比报告 → — V1 Direct / V2 Direct / Claude Code / Codex 在真实数据分析场景下的全维度评估。

A local-first, multi-engine AI agent dashboard. Run Claude Code, Codex, and a built-in ReAct loop side-by-side from one window. Local SQLite history + long-term memory that extracts durable facts automatically. No account, no relay server — your LLM API key is the only auth.

English | 简体中文

Why KinetAios?

Most AI clients lock you into one provider, lose context when you switch engines, and route your conversations through a relay server. KinetAios runs three engines from one window, with cross-engine long-term memory and no account.

Install

Download the latest release:

• Windows — KinetAios-Setup-1.0.0.exe (NSIS installer)
• macOS — see releases

Unsigned build → Windows SmartScreen / macOS Gatekeeper will warn; allow manually.

First launch: click ⚙ top-right → fill in API Key (+ Base URL / model; default GLM Zhipu) → once "Test connection" passes, send a task.

Run from source

Requires Node.js 18+ and internet (the better-sqlite3 native module needs to compile).

cd KinetAiosWin
npm install      # postinstall rebuilds better-sqlite3 for Electron
npm run build
npm start

On a CN network npm install may time out fetching the Electron binary — .npmrc is already configured with the npmmirror mirror; on failure you can also run ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/ node node_modules/electron/install.js.

Features

Three engines (switchable per session; switching clears cross-engine context)

• Direct (Kaios): built-in ReAct loop + GLM/OpenAI-compatible & Anthropic dual-protocol SSE streaming provider, with tool-level concurrency, sub-agent dispatch, context compaction, and retry.
• Claude Code: spawns claude -p --output-format stream-json, parses NDJSON, resumes via --resume.
• Codex: spawns codex exec --json, parses JSONL, resume continuation.

Direct tools (12)

shell (confirm before exec), read_file, write_file, edit_file (precise replacement), grep (recursive content search), glob (list files), web_fetch (SSRF-protected, Jina Reader fallback), web_search (Bing → DuckDuckGo), recall_memory, git_diff (read-only), dispatch_agent (read-only sub-agent with independent context), flight_plan (plugin-injected tools may add more).

MCP integration (client + server)

• Client: auto-discovers system-configured MCP services (~/.claude.json, ~/.codex/config.toml, Claude Desktop config), stdio transport, auto-reconnect. 🔌 button shows connected services/tools.
• Server: built-in MCP Server (HTTP+SSE) exposes run_agent — remote machines can invoke your local agent with full tool access. Token-authenticated (timing-safe comparison), 5-min timeout, zombie-connection detection.

Long-term memory + memory graph

• Auto-extracts durable facts from each turn → SQLite → injected into next turn. Cross-engine, cross-session.
• Memory graph visualization: force-directed graph showing memory provenance, conflict detection, and timeline. Separate full-screen window.
• Import/export memories as JSON (migration / backup).

Skills / Commands / Agents / Plugins

• Scans Claude Code's skills + commands + agents and Codex's skills. / menu or ⚡ button.
• Plugin system v2.2: plugins can contribute tools, slash commands, hooks, and full-screen panels. Per-need injection (keyword matching saves ~60% tokens). Built-in plugins: office-suite, brainstorm (Excalidraw), math-practice, cpp-learning, low-altitude, and more.

Sidebar (left → right)

• ＋ New session.
• 📂 Workbench — project cards grouped by cwd, showing recent activity + cost. "Context" button edits KINET-CONTEXT.md.
• 📊 Dashboard — independent window: real-time token usage, cost stats, engine distribution (all sessions aggregated).
• 🌐 Files — file browser + <webview> preview (HTML/SVG/PNG/JPG/PDF) + textarea editor. Multi-tab support. Address bar accepts file:// / http(s):// / localhost:<port>.
• 🏘️ Town — game-style isometric visualization of remote nodes (other KinetAios instances on your network).
• 🧠 Memory — memory panel: current channel / all scope, inline edit/delete, provenance.
• 🔌 Plugins — plugin management: enable/disable, search, category cards.
• ⚙️ Settings — see below.

Main window tabs (Chat / Files / Git / Rules)

• Chat — streaming output, collapsible tool steps, real-time token count, context inspector, screenshot, voice input.
• Files — same as 🌐 sidebar, follows current session cwd.
• Git — git status (left) + git log (right). Click file → side-by-side diff; click commit → formatted git show.
• Rules — edit cwd's KinetAios.md (project-level rules, injected into system prompt).

Pipeline (cross-engine orchestration)

Chain multiple stages, each specifying an engine + prompt. Previous stage's output auto-prepends to next stage's prompt. 2-min per-stage timeout with polling, fail-fast abort.

Session branching & handoff

• Branch: fork from any turn — deep-copies turns/steps into a new session.
• Export/import session: serialize full session state (turns + history + engine + model + cwd) for cross-machine handoff. Sensitive data (API keys, secrets) auto-redacted on export.
• Cross-session references: link related sessions, visualized as a DAG task graph.

Context management (Direct engine)

• Context inspector: view/edit the raw directHistory array per session (JSON textarea editor).
• Auto-compaction: when history exceeds token budget, early turns are summarized by the LLM. Compaction events visualized in the UI (before/after token counts).
• Per-protocol token calibration: token/char ratio tracked separately per API protocol (OpenAI vs Anthropic), preventing concurrent sessions from skewing each other's estimates.

Multimodal (Direct engine)

• Image input: 📎 select/paste images → vision content parts → OpenAI image_url or Anthropic base64 format.
• Voice: 🎤 record → Whisper transcription → fills composer. TTS via speechSynthesis (auto language detection, 2000-char limit).
• Screenshot: 📸 overlay → drag-select region → cropped image injected into prompt.

Global search

Overlay (Ctrl/Cmd+K) searches across all conversations — matches prompt text, answer text, and tool output.

Settings (⚙️)

• API: provider (OpenAI / Anthropic), base URL, model, key. GLM / DeepSeek / OpenAI / Anthropic presets. Balance check button (GLM Zhipu). Encrypted via safeStorage.
• Behavior: shell approval mode, sandbox level, plan mode, CLI engine toggle, close behavior (quit / minimize / tray).
• Pricing: per-model input/output prices for cost calculation.
• Interface: language (English / 简体中文 / 繁體中文 / 日本語), theme (dark / light, live preview).
• Memory: export/import JSON.

Other

• Per-session model (editable dropdown, OpenAI-compatible + Anthropic dual-protocol).
• File attachments: 📎 select/drag multiple text files (large files truncated), @path references.
• KinetAios.md / AGENTS.md / CLAUDE.md: project rules auto-injected into system prompt.
• Tray + global hotkey Ctrl/Cmd+Alt+Space → quick panel.
• Configurable brand (brand.json), encrypted API key storage (safeStorage: macOS Keychain / Windows DPAPI).

Tech stack

• Electron + TypeScript — the main process runs the agent runtime; the renderer is native web UI.
• better-sqlite3 — SQLite + FTS5 (history / recall_memory full-text search + embedding-based semantic recall).
• No frontend framework — renderer is pure vanilla TS + HTML/CSS, bundled with esbuild.

Directory layout

KinetAiosWin/
  brand.json               # brand config (product name etc., read at startup)
  package.json
  src/
    shared/types.ts         # types + applyEvent (shared by main/renderer, single source of truth)
    shared/i18n.ts          # four-language string table + t()
    main/
      main.ts               # windows / tray / hotkey / IPC / shell-confirm bridge
      TaskManager.ts        # session management + engine dispatch + memory extraction
      engines.ts            # Engine interface + Direct/ClaudeCode/Codex + cross-platform CLI spawn
      AgentLoop.ts          # ReAct loop (Direct) + history compaction + reactive trim
      glm.ts                # Provider + OpenAI/Anthropic SSE streaming + retry
      tools.ts              # 12 built-in tools + cross-platform shell + dispatch_agent
      mcp.ts                # MCP client (scan + stdio + reconnect)
      mcp-server.ts         # MCP server (HTTP+SSE, run_agent, token auth)
      skills.ts             # skills/commands/agents/plugin scan
      plugins.ts            # plugin loader (v2.2: tools/slashCommands/hooks/panels)
      store.ts              # better-sqlite3 + FTS5
      settings.ts           # config (encrypted API key persistence, lang, embeddings)
    preload/preload.ts      # the narrow API exposed via contextBridge
    renderer/
      index.html quick.html styles.css
      app.ts                # dashboard logic (chat, sidebar, tabs, context inspector)
      quick.ts              # quick panel logic
      dashboard.ts          # cost/token dashboard window
      arena.ts              # deep analytics dashboard
      memory-graph.ts       # memory graph SVG visualization
      town.ts               # Town view (remote node visualization)
      files-pane.ts         # file browser + webview preview + editor
      markdown.ts           # mini markdown renderer

Build / dev

npm run build       # tsc (main) + esbuild (renderer) + copy brand.json
npm run typecheck   # typecheck both halves (no output)
npm start           # launch (requires a prior build)
npm run dev         # build + start

Package

npm run dist         # current platform's default target

• Windows — release\KinetAios Setup <ver>.exe (NSIS). Must be built on Windows (cross-building Windows + native modules from macOS is unreliable).
• macOS — build a dmg: npx electron-builder --mac (needs a mac toolchain).
• electron-builder rebuilds better-sqlite3 against Electron's ABI; asar: false avoids native-module load errors from inside an asar.
• Unsigned → Windows SmartScreen / macOS Gatekeeper will warn; users allow manually. Removing the warning needs a signing cert + Apple notarization.
• The default icon is Electron's; to use your own: Windows build/icon.ico (256×256), mac build/icon.icns.

Known constraints

• Close behavior is configurable (quit / minimize / tray); default is minimize. The global hotkey only works while the app runs.
• Mac→Windows cross-build of the native module is unreliable — build Windows installers on a Windows machine or via windows-latest GitHub Actions.