agentweave-ai 0.18.0

Multi-agent AI collaboration framework (Claude, Kimi, Gemini, Codex, and more)

AgentWeave

A collaboration framework for N AI agents — Claude, Kimi, Gemini, Codex, Minimax, GLM, and more

📖 Documentation

AgentWeave lets multiple AI agents work together on the same project through a shared protocol. The AgentWeave Hub is a self-hosted server with a web dashboard — the recommended way to run it.

---

Quick Start — AgentWeave Hub (Recommended)

The Hub provides a web dashboard, REST + SSE + MCP interfaces, and real-time visibility into agent activity.

Step 1 — Start the Hub (Docker)

# Download the two config files
curl -O https://raw.githubusercontent.com/gutohuida/AgentWeave/master/hub/docker-compose.yml
curl -O https://raw.githubusercontent.com/gutohuida/AgentWeave/master/hub/.env.example

# Create your .env
cp .env.example .env

Open .env and set your API key:

# Generate a secure key
python -c "import secrets; print('aw_live_' + secrets.token_hex(16))"

Paste the output as AW_BOOTSTRAP_API_KEY in .env, then start the Hub:

docker compose up -d

The Hub is now running at http://localhost:8000 — open it in your browser to see the dashboard.

---

Step 2 — Install the CLI

pip install "agentweave-ai[mcp]"

---

Step 3 — Initialize your project

cd /path/to/your-project
agentweave init --project "My App" --agents claude,kimi

This creates AI_CONTEXT.md (fill it in once: stack, architecture, standards) and .agentweave/ with agent roles and shared context.

---

Step 4 — Connect the CLI to the Hub

agentweave transport setup --type http \
  --url http://localhost:8000 \
  --api-key aw_live_<your-key> \
  --project-id proj-default

---

Step 5 — Register the MCP server and start the watchdog

# Register MCP with all session agents (one command)
agentweave mcp setup

# Start the background watchdog (one terminal, all agents)
agentweave start
# Stop later with: agentweave stop

Restart your Claude / Kimi sessions so they pick up the new MCP server. That's it — agents communicate through the Hub and you monitor everything in the dashboard.

---

What the Dashboard Shows

Open http://localhost:8000 to see:

• Tasks board — all tasks with status, priority, assignee, requirements, acceptance criteria, and deliverables (click any card to expand)
• Messages feed — inter-agent messages with expand-to-read for long content; message type and linked task shown inline
• Human questions — questions agents have asked you; answer directly in the dashboard
• Agent activity — live event stream and per-agent output log
• Agent cards — connected agents auto-discovered from your session; shows role, yolo mode, and per-agent chat history

---

Configuration — .env reference

Variable	Default	Description
AW_BOOTSTRAP_API_KEY	(required)	API key auto-created on first start (aw_live_…)
AW_BOOTSTRAP_PROJECT_ID	proj-default	Default project ID
AW_BOOTSTRAP_PROJECT_NAME	Default Project	Display name for the default project
AW_PORT	8000	Port the Hub listens on
AW_CORS_ORIGINS	(empty)	Comma-separated allowed origins for CORS (leave empty in production)
DATABASE_URL	sqlite+aiosqlite:///data/agentweave.db	SQLite path inside the container

Data persists in a Docker volume (hub-data) — no manual backup needed for local use.

---

Alternative Modes

Mode	Setup	Best for
Hub	Docker + agentweave transport setup --type http	Teams, multi-machine, web dashboard (recommended)
Zero-relay MCP	agentweave mcp setup + watchdog	Autonomous loops, same machine, no server
Manual relay	Zero — just install	Quick one-off delegation

Zero-relay MCP (no Hub)

pip install "agentweave-ai[mcp]"
cd your-project/
agentweave init --project "My App" --agents claude,kimi
agentweave mcp setup   # configure MCP in agent settings
agentweave start       # start background watchdog

Manual relay (simplest possible)

pip install agentweave-ai
cd your-project/
agentweave init --project "My App" --agents claude,kimi
# Ask Claude to delegate; it runs agentweave quick + relay and gives you a prompt to paste into Kimi

---

Claude-Proxy Agents (Minimax, GLM, and any OpenAI-compatible provider)

Some models — like MiniMax and Zhipu GLM — don't have a native CLI. AgentWeave runs them through the Claude Code CLI by overriding two environment variables:

ANTHROPIC_BASE_URL  →  the provider's OpenAI-compatible endpoint
ANTHROPIC_API_KEY   →  the provider's API key (resolved from your shell at runtime)

This is called a claude_proxy runner. AgentWeave tracks a separate Claude resume session ID per proxy agent so each one maintains its own conversation history.

Setup

# 1. Initialize with a proxy agent
agentweave init --project "My App" --agents claude,minimax

# 2. Configure the runner (uses built-in defaults for minimax and glm)
agentweave agent configure minimax

# Or supply custom values for any OpenAI-compatible provider:
agentweave agent configure mymodel \
  --runner claude_proxy \
  --base-url https://api.example.com/v1 \
  --api-key-var MY_MODEL_API_KEY

Running a proxy agent

# Option A — switch env vars in your current shell, then run Claude manually
export MINIMAX_API_KEY=<your-key>
eval $(agentweave switch minimax)          # exports ANTHROPIC_BASE_URL + ANTHROPIC_API_KEY
claude --resume <session-id> -p "..."

# Option B — let AgentWeave handle it (sets env + launches Claude with relay prompt)
export MINIMAX_API_KEY=<your-key>
agentweave run --agent minimax

# Option C — from relay (shows switching instructions, or auto-runs with --run)
agentweave relay --agent minimax           # shows copy-paste prompt + switching instructions
agentweave relay --agent minimax --run     # combined: sets env + launches Claude

Session continuity

AgentWeave tracks the Claude session ID per proxy agent so --resume is used automatically on subsequent runs. To register a session ID manually (e.g. from claude --list):

agentweave agent set-session minimax <session-id>

Security note: API keys are never stored in session.json. Only the env var name is stored (e.g. MINIMAX_API_KEY). The actual value is resolved from your shell at runtime.

Built-in provider registry

Agent	Base URL	Env var
minimax	https://api.minimax.chat/v1	MINIMAX_API_KEY
glm	https://open.bigmodel.cn/api/paas/v4	ZHIPU_API_KEY

---

Cross-Machine Collaboration

Via Git (no server required)

agentweave transport setup --type git --cluster yourname

Creates an orphan branch (agentweave/collab) on your git remote. Messages sync through git plumbing — working tree and HEAD are never touched. Both developers need access to the same remote.

Via Hub (recommended for teams)

Deploy the Hub once, connect all agents via HTTP transport. The dashboard shows all messages, tasks, and human questions in real time.

---

Commands Reference

Session

agentweave init --project "Name" --agents claude,kimi
agentweave status
agentweave summary

Delegation

agentweave quick --to kimi "Task description"
agentweave relay --agent kimi
agentweave relay --agent minimax --run     # auto-run for claude_proxy agents
agentweave inbox --agent claude

Agent runner (claude_proxy setup)

agentweave agent configure minimax                      # use built-in defaults
agentweave agent configure glm                          # use built-in defaults
agentweave agent configure mymodel \                    # custom OpenAI-compatible provider
  --runner claude_proxy \
  --base-url https://api.example.com/v1 \
  --api-key-var MY_MODEL_API_KEY
agentweave agent set-session minimax <session-id>       # register Claude resume ID manually

agentweave switch minimax        # output eval-able export commands
agentweave run --agent minimax   # set env vars + launch Claude with relay prompt

Tasks

agentweave task list
agentweave task show <task_id>
agentweave task update <task_id> --status in_progress
agentweave task update <task_id> --status completed
agentweave task update <task_id> --status approved
agentweave task update <task_id> --status revision_needed --note "Fix X"

Transport

agentweave transport setup --type http --url ... --api-key ... --project-id ...
agentweave transport setup --type git --cluster yourname
agentweave transport status
agentweave transport pull
agentweave transport disable

Yolo mode

agentweave yolo --agent claude --enable    # allow agent to act without confirmations
agentweave yolo --agent claude --disable   # re-enable confirmation prompts

Human interaction (Hub only)

agentweave reply --id <question_id> "Your answer"

---

MCP Tools Reference

Available to agents in both local MCP mode and via Hub MCP:

Tool	What it does
send_message(from, to, subject, content)	Send a message to another agent
get_inbox(agent)	Read unread messages
mark_read(message_id)	Archive a message after processing
list_tasks(agent?)	List active tasks
get_task(task_id)	Get full task details
update_task(task_id, status)	Update task status
create_task(title, ...)	Create and assign a new task
get_status()	Session-wide summary + task counts
ask_user(from_agent, question)	Post a question to the human (Hub only)
get_answer(question_id)	Check if the human answered (Hub only)

---

Task Status Lifecycle

pending → assigned → in_progress → completed → under_review → approved
                                             ↘ revision_needed (loops back)
                                             ↘ rejected

---

Build from Source

git clone https://github.com/gutohuida/AgentWeave.git
cd AgentWeave/hub

cp .env.example .env
# Edit .env: set AW_BOOTSTRAP_API_KEY

docker compose up --build -d

Hub UI development (hot-reload)

cd hub/ui
npm install
npm run dev      # dashboard at http://localhost:5173, proxies /api → Hub at localhost:8000

---

Repository Layout

AgentWeave/
├── src/agentweave/     CLI package (Python 3.8+, zero runtime deps) — v0.18.0
├── hub/                AgentWeave Hub server (Python 3.11+, FastAPI + Docker) — v0.12.0
│   ├── hub/            Hub Python package
│   ├── ui/             React dashboard (built into Docker image, no separate server)
│   └── Dockerfile      Multi-stage build: Node UI → Python server
├── docs/               Additional documentation
├── tests/              CLI unit tests (pytest)
└── Makefile            Convenience targets for both packages

---

Development

# CLI
pip install -e ".[dev]"
ruff check src/
black src/
mypy src/
pytest tests/ -v

# Hub
cd hub
pip install -e ".[dev]"
make ui-build    # rebuild React UI
pytest tests/ -v

# Both
make install-all
make test-all
make lint

---

Roadmap

Phase	Status	Description
Local transport	✅ Done	Single-machine via .agentweave/ filesystem
Git transport	✅ Done (v0.2.0)	Cross-machine via orphan branch, zero infra
N-agent support	✅ Done (v0.3.0)	Multi-agent teams with roles.json and cluster naming
Local MCP server	✅ Done (v0.4.0)	Native tool integration, zero-relay with watchdog pinger
HTTP transport	✅ Done (v0.5.0)	CLI ↔ Hub via REST
AgentWeave Hub	✅ Done (v0.2.0)	Self-hosted server, REST + SSE + MCP + web dashboard
Hub UI	✅ Done (v0.2.1)	React dashboard — expandable tasks/messages, agent trigger, configurator
Per-agent context templates	✅ Done (v0.6.0)	claude_context.md, kimi_context.md, collab_protocol.md
Session sync to Hub	✅ Done (v0.10.0)	Watchdog pushes session.json to Hub on startup; agents auto-appear in dashboard
Yolo mode	✅ Done (v0.10.0)	Per-agent flag to suppress confirmation prompts for autonomous loops
Claude-proxy agents	✅ Done (v0.12.0)	Run Minimax, GLM, and any OpenAI-compatible provider via Claude CLI proxy
Multi-role support	✅ Done (v0.15.0)	Multiple roles per agent with agentweave roles CLI and Hub sync
Official hosted Hub	🔲 Planned	Public hub.agentweave.dev — Supabase + Vercel + Railway

---

FAQ

Q: Do I need the Hub? No. Manual relay and local MCP modes work with zero infra. The Hub adds a web dashboard, multi-machine support, and human question-answering.

Q: Should I put the UI in a separate folder/repo? No. The UI (hub/ui/) is built into the Docker image and served by the Hub at the same port. No second server or CORS config needed in production.

Q: Do I need to run CLI commands during my session? No. After agentweave init, just talk to Claude. It runs all agentweave commands via Bash automatically.

Q: Do the watchdog processes need to stay running? Yes (in local MCP mode or Hub mode). Run agentweave start once. If they stop, messages still queue — agents just won't be auto-triggered.

Q: Should I commit .agentweave/? Partially. Runtime state (tasks, messages, session.json, transport.json) is gitignored. AGENTS.md and README.md are committed.

Q: Do both developers need the same git remote for git transport? Yes. Git transport requires a shared remote (e.g. origin).

Q: How do I use Minimax or GLM — they don't have a CLI? Use agentweave agent configure minimax (or glm) to set them up as claude_proxy agents. AgentWeave runs them through the Claude CLI with overridden env vars (ANTHROPIC_BASE_URL + ANTHROPIC_API_KEY). Export your provider key, then run agentweave run --agent minimax. See the Claude-Proxy Agents section above.

Q: Can I use any OpenAI-compatible provider as an agent? Yes. Use agentweave agent configure <name> --runner claude_proxy --base-url <url> --api-key-var <VAR>. The Claude CLI will proxy requests to that endpoint using your existing API key env var.

---

Links

• GitHub: https://github.com/gutohuida/AgentWeave
• PyPI: https://pypi.org/project/agentweave-ai/
• Issues: https://github.com/gutohuida/AgentWeave/issues
• Roadmap: ROADMAP.md

---

MIT License