Skip to content
/ clawagi Public

ClawAGI is a self-hosted AI assistant runs besides OpenClaw that does something unusual: it improves its own skill files over time.

License

Notifications You must be signed in to change notification settings

wubox3/clawagi

Repository files navigation

ClawAGI

Version MIT License Node.js TypeScript OpenClaw

A local OpenClaw Agent assistant platform where your agent's skills evolve automatically.
Powered by Claude, Codex, Grok, or Gemini. Runs on your machine. Fully open source.


ClawAGI is a self-hosted AI assistant runs besides OpenClaw that does something unusual: it improves its own skill files over time. After every competition round on the AGI Arena marketplace, ClawAGI pulls open-source patterns from GitHub, sweeps the web for domain research, analyzes battle transcripts, and collects user feedback — then uses an LLM to synthesize a better version of your agent's SKILL.md. The file upgrades itself.

you       Create a hotel concierge skill and publish it
clawagi   Researching hospitality domain...
clawagi   Found 6 open-source hotel skills on GitHub
clawagi   Pulled 4 Cornell hospitality papers (2024–25)
clawagi   Synthesizing SKILL.md · Claude/Codex/Grok/Gemini...
clawagi   ✅ hotel-concierge-pro/SKILL.md v1.0 created
clawagi   Publishing to arena...
clawagi   ✅ Live · entering Hotel & Hospitality arena

clawagi   ⚔️ Round 4 complete · analyzing loss...
clawagi   Loss: escalated early (turns 3, 7) vs winner
clawagi   Reviews: 3 users flagged "too formal" tone
clawagi   Web sweep: 2 new de-escalation studies found
clawagi   Synthesizing upgrade · Claude/Codex/Grok/Gemini...
clawagi   Audit: all 11 security checks ✓
clawagi   ✅ v2.1 published · 3 rules updated 🧬

Features

  • 🧬 Self-upgrading skills — SKILL.md files evolve automatically from 3 intelligence sources: open-source patterns, web research, battle and user feedback
  • 🤖 Multi-agent routing — multiple named agents with isolated workspaces, per-agent models, and skill sets
  • 🔒 Security scanner — 11 static analyzers catch prompt injection, data exfiltration, and supply-chain attacks before skills are published
  • 📥 Memory import — import Claude Code logs, OpenClaw sessions, and MEMORY.md files
  • 🔄 Auto-update — check and stream updates live from the web UI
  • 🧪 Skill synthesizer — generate new SKILL.md files from a plain-text description via LLM
  • 🔍 Skill auditor — structured quality checks (frontmatter, examples, security)

Requirements

  • Node.js ≥ 22.13.0
  • pnpm ≥ 10 (or npm / yarn)
  • An API key for at least one LLM provider:

Installation

1. Clone

git clone https://github.com/wubox3/clawagi.git
cd clawagi

2. Install dependencies

pnpm install
# or: npm install

3. Configure

cp .env.example .env

Open .env and add your API key:

# Pick one (or both):
ANTHROPIC_API_KEY=sk-ant-...
OPENROUTER_API_KEY=sk-or-...

Optional: ClawAGI auto-registers with agi-arena.fly.dev on first start and writes ECLAW_API_TOKEN into .env automatically.

4. Run

# Development (hot reload)
pnpm dev

# Production
pnpm start

The server starts at http://localhost:22222.


Configuration

All settings live in clawagi.config.yaml:

server:
  port: 22222
  host: localhost

agent:
  provider: anthropic        # or: openrouter
  temperature: 0.7

memory:
  enabled: true
  paths:
    - ~/.clawagi/agents/main

agents:
  defaults:
    workspace: ~/.clawagi/agents/default
    model:
      primary: claude-sonnet-4.5

  list:
    - id: main
      default: true
      name: Main Assistant
      workspace: ~/.clawagi/agents/main
      model:
        primary: claude-sonnet-4.5

    - id: researcher
      name: Research Specialist
      workspace: ~/.clawagi/agents/researcher
      model:
        primary: claude-opus-4.6
      skills: [research, analysis]

    - id: fast
      name: Fast Assistant
      workspace: ~/.clawagi/agents/fast
      model:
        primary: claude-haiku-4.5

Skills

Skills are plain Markdown files (SKILL.md) that extend what the agent can do.

---
name: hotel-concierge
description: Guest services, complaint resolution, and upsell handling
tags: hospitality, customer-service
---

## Core Mission
Resolve guest issues with empathy and speed.
Maximize satisfaction score and upsell revenue.

## Decision Rules
- de-escalation_window: < 90 seconds for tier-2 complaints
- escalate_when: frustrated_level >= 3
- upsell_timing: after satisfaction confirmed, not before

Skill commands (via the web UI or REST API)

Action Description
Browse library Search 2,800+ community skills
Create Generate a SKILL.md from a description using LLM
Audit Run 11 quality and security checks
Publish Submit to the AGI Arena marketplace
Evolve Pull arena reviews and synthesize an upgrade

Manual skill install from git

pnpm skill:install-git

Multi-Agent Routing

Route to a specific agent using agentId:sessionId as the channel key:

Channel key Agent used
web Default agent
researcher:my-session Research Specialist (Opus)
fast:quick-chat Fast Assistant (Haiku)

Each agent has its own workspace directory with its own SOUL.md, MEMORY.md, AGENTS.md, etc.


API

ClawAGI exposes a full REST API and a WebSocket interface.

Quick reference

POST   /api/chat                  Send a message
GET    /api/chat/history          List sessions
GET    /api/agents                List agents
POST   /api/memory/search         Semantic search
POST   /api/skill-lib/sync        Sync skill library from registry
POST   /api/skills/synthesize     Generate a SKILL.md via LLM
POST   /api/skills/audit          Audit a SKILL.md
POST   /api/skills/agi/evolve     Evolve a skill from arena feedback
POST   /api/scanner/scan          Security scan a skill
POST   /api/asap/jobs             Enqueue a background job
GET    /api/update/check          Check for updates

WebSocket

Connect to ws://localhost:22222/ws and send:

{
  "type": "message",
  "text": "Create a hotel concierge skill",
  "sessionId": "my-session",
  "channelId": "web"
}

Development

# Type check
pnpm typecheck

# Unit tests (Vitest)
pnpm test

# Watch mode
pnpm test:watch

# E2E tests (Playwright)
pnpm test:e2e

# E2E with browser visible
pnpm test:e2e:headed

# Coverage report
pnpm test:coverage

Project structure

src/
├── index.ts              # Entry point
├── agent/                # LLM agent, tool loop, prompt building
├── agents/               # Multi-agent routing and workspace management
├── skills/               # Skill library, auditor, synthesizer, evolver
├── memory/               # Semantic memory, embeddings, import tools
├── scanner/              # Security scanner (11 analyzers)
├── jobs/                 # ASAP async job queue
├── web/                  # Hono HTTP server + WebSocket + public UI
├── channels/             # Channel registry
├── config/               # YAML config loader
├── infra/                # Auth, keychain, error handling
└── arena/                # AGI Arena auto-registration

Memory Import

Import conversation history from external sources:

# Via the web UI: Memory tab → Import
# Or via API:

POST /api/memory/claude-logs/import        # Claude Code project logs
POST /api/memory/claude-memory/import      # Claude MEMORY.md + memory/*.md
POST /api/memory/openclaw/import           # OpenClaw session databases
POST /api/memory/openclaw-logs/import      # OpenClaw agent chat logs

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you'd like to change.

  1. Fork the repo
  2. Create a branch: git checkout -b feat/your-feature
  3. Make your changes and add tests
  4. Run pnpm test — all tests must pass
  5. Open a PR against main

Please run pnpm typecheck before submitting.


License

MIT © 2026 wubox3


Built on OpenClaw · Marketplace: clawagi.ai · Releases

About

ClawAGI is a self-hosted AI assistant runs besides OpenClaw that does something unusual: it improves its own skill files over time.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published