Agim · Universal IM-to-Agent Bridge

中文说明

Agim connects IM channels to AI agents with one Node.js process. It is designed for production use: at-least-once persistent delivery, session continuity, human approval, and observable operations.

What matters

• Bridge first: WeChat / Feishu / DingTalk / Email to multiple agent backends. Telegram / Discord stay hidden unless AGIM_ENABLE_GLOBAL_IM=1.
• Agent flexibility: Claude Code, Codex, GitHub Copilot CLI, OpenCode, Cursor, Kimi Code, Qoder, PI, MiMo Code, Antigravity, and Agim Agent (in-process LLM).
• AI workflows built-in: memory, skills, Focus, Loop Engineering, Subagents (call_agent / call_agents), heartbeat, reminders, memos, ask-user.
• Reliable runtime: SQLite outbox (at-least-once delivery), restart recovery, job board, structured audit. Production requires native better-sqlite3; without it the outbox is unavailable.
• Safety by default: token auth, sensitive-path denylist. Tool auto-approve is off by default; opt in with AGIM_AUTO_APPROVE=1 (and a strict sender allowlist when IM is enabled).

Quick start

Requires Node.js ≥ 22.19. A C++ toolchain (gcc/g++, make, python3) is needed so better-sqlite3 can compile on install.

After install, run agim (or agim go). The first command starts the service and opens the browser wizard: pick a model, paste one API key, then chat.

npm install -g agim-cli
agim

First run will:

1. Detect usable local agents such as Claude Code, Codex, GitHub Copilot CLI, Cursor, Kimi Code, Qoder, PI, MiMo Code, Antigravity, or Agim Agent.
2. If none are ready, start Agim and open http://localhost:3000/setup so you can pick DeepSeek / Kimi / GLM / MiMo / Ollama and paste one API key.
3. Print the Web console URL. On first boot a one-time token is written to ~/.agim/web-bootstrap-token (retrieve with agim token bootstrap if needed). Loopback (127.0.0.1) usually skips login.

Default Web console: http://localhost:3000 · first-run wizard: /setup
Skip writes setupSkipped and leaves /setup available. Details: docs/first-run-setup.md.

• Chat: /
• Mobile Portal: /m/chat (/ Skills·Commands, @ workspace refs, Focus / Loop docks — see docs/portal-composer.md)
• Native shells: iOS apps/ios · macOS apps/macos (both wrap /m/chat with host + web session cookie; see docs/macos-shell.md)
• Tasks: /tasks (Loop audit at /tasks/loops)
• Settings: /settings

If you forget the URL or token path:

agim web

Local health check:

agim doctor

After the Web chat works, connect WeChat / Feishu / DingTalk from Settings. agim quickstart is kept as a compatibility alias.

Common CLI entries:

Choose your AI backend

Agim Agent setup details: docs/llm-backends.md (DeepSeek V4: §8.1 compat, §8.2 API protocol / beta, §8.3 RPC).

Supported agent details: docs/agents.md

Core AI workflows

Security defaults

• Web token auth enabled by default.
• Auto-approve disabled by default (AGIM_AUTO_APPROVE=0): mutating tools go through IM/Web approval cards. Set AGIM_AUTO_APPROVE=1 for single-operator convenience; with IM messengers configured, also set a strict AGIM_ALLOWED_USERS=platform:userId,... (or acknowledge open IM with AGIM_AUTO_APPROVE_OPEN_IM=1).
• Sensitive files/paths are denied by policy (hard-deny, cannot be overridden).
• Self-protect prevents agents from killing/restarting agim.
• Chat /restart is IM-admin only. Web console and Portal /m/chat refuse it; use Settings → Service → Restart or agim restart (docs/runbook.md §4).
• Telegram / Discord require AGIM_ENABLE_GLOBAL_IM=1 (domestic defaults keep them off).

Security reference: docs/security-hardening.md

Essential chat commands

Type /help in chat for the full list.

Minimal config shape

~/.agim/config.json

{
  "messengers": [],
  "agents": ["agim", "claude-code"],
  "defaultAgent": "agim",
  "llmBackends": [
    {
      "name": "default-agim",
      "provider": "openai-compat",
      "baseUrl": "https://api.openai.com/v1",
      "model": "gpt-4o-mini"
    }
  ],
  "llmRoles": {
    "cheap": "default-agim",
    "finance": "default-agim"
  }
}

IM is optional — leave messengers empty for web-console-only, then add WeChat / Feishu / DingTalk from Settings or agim config.

Secrets are stored separately from this file. Task dispatch (coding / finance / collect) can be configured via dropdowns under Settings → LLM → Task dispatch cheat sheet (see docs/llm-backends.md §6.1).

Operations

• Deployment guide: docs/deployment.md
• Runbook: docs/runbook.md
• Architecture stance: docs/architecture-bridge-and-agim-agent.md
• LLM backends & dispatch cheat sheet: docs/llm-backends.md
• Local Unix RPC (opt-in llm_complete / static token for same-host services): see CHANGELOG.md Unreleased
• Full changelog: CHANGELOG.md

Docs index

Start from: docs/README.md

Migration notes

Current public surfaces use the Agim naming consistently:

• executable: agim
• environment variables: AGIM_*
• data paths: ~/.agim and ~/.agim-workspaces

/goal is retired — use Focus (/focus) and Loop (/loop) instead (docs/goals.md, docs/architecture/focus.md).

Development

git clone https://github.com/benking007/agim.git
cd agim
npm ci
npm --prefix src/web-app ci
npm run build
npm run lint
npm run typecheck
npm test

License

MIT