agentweave-ai 0.4.0

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

AgentWeave

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

AgentWeave lets multiple AI agents work together on the same project. Agents communicate through a shared .agentweave/ directory and a local MCP server. With MCP enabled, agents call send_message and get_inbox as native tools — no manual relay needed.

---

Two Modes

Mode 1 — Manual relay (zero extra setup)

The original mode. You paste a relay prompt from Claude into Kimi, and back. Works with no dependencies beyond the base package.

Mode 2 — Zero-relay with MCP (recommended)

Install the MCP extras, run two background watchers, and agents communicate autonomously. The loop:

Claude sends message via send_message tool
  → watchdog detects it
  → fires: kimi --print -p "check your inbox"
  → Kimi calls get_inbox, reads it, replies via send_message
  → watchdog detects it
  → fires: claude -p "check your inbox"
  → Claude reads and continues

You only interact with Claude — the human never relays anything.

---

Quick Start

1. Install

# Base package (manual relay mode)
pip install agentweave-ai

# With MCP server (zero-relay mode)
pip install "agentweave-ai[mcp]"

2. Initialize (once per project)

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

Creates:

• AI_CONTEXT.md (project root) — project DNA: stack, architecture, code standards (static, rarely changes)
• .agentweave/AGENTS.md — collaboration protocol: MCP vs manual mode, workflow (per-session)
• .agentweave/ROLES.md — agent role assignments: who does what (per-session)
• .agentweave/shared/context.md — current focus: what's being worked on today (changes daily)

Supported agents: claude, kimi, gemini, codex, aider, cline, cursor, windsurf, copilot, and any name matching ^[a-zA-Z0-9_-]{1,32}$

3. Fill in project context

AI_CONTEXT.md — Fill this once with your project's fundamentals:

• Project overview, tech stack, essential commands
• Architecture, directory structure
• Code standards and workflow conventions

shared/context.md — Update this daily with current state:

• What phase you're in, what's being worked on
• Recent decisions, blockers, next steps

Agents read all four files on session start.

4a. Start working — manual relay mode

Just prompt Claude. It runs all agentweave CLI commands via Bash automatically:

"Claude, delegate the database schema design to Kimi."

Claude runs agentweave quick and agentweave relay, then shows you a prompt to paste into Kimi Code. When Kimi is done, tell Claude and it reads the reply.

4b. Start working — zero-relay MCP mode

One-time MCP setup:

# Configure the MCP server in both agents (auto-detects claude and kimi CLI)
agentweave mcp setup

Or manually:

claude mcp add agentweave -- agentweave-mcp
kimi mcp add --transport stdio agentweave -- agentweave-mcp

Start the watchers (two terminals, keep them running):

# Terminal 1 — notifies Claude when Kimi sends a message
agentweave-watch --auto-ping --agent claude

# Terminal 2 — notifies Kimi when Claude sends a message
agentweave-watch --auto-ping --agent kimi

Now just prompt Claude. It uses send_message and get_task as native MCP tools. When it sends a message to Kimi, the watchdog fires kimi --print -p "check inbox" automatically, and the loop continues without you.

---

MCP Tools Reference

Once configured, both agents have these tools available natively:

Tool	What it does
send_message(from_agent, to_agent, 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, optionally filtered
get_task(task_id)	Get full task details
update_task(task_id, status)	Update task status
create_task(title, description, assignee, ...)	Create and assign a new task
get_status()	Session summary + task counts

---

Commands Reference

Session

agentweave init --project "Name" --principal claude   # Initialize
agentweave status                                      # Full status
agentweave summary                                     # Quick overview

Delegation (manual relay mode)

agentweave quick --to kimi "Task description"         # Create + assign task
agentweave relay --agent kimi                         # Generate relay prompt

Tasks

agentweave task list                                  # List all tasks
agentweave task show <task_id>                        # View task details
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"

Messaging

agentweave inbox --agent claude                       # Check Claude's inbox
agentweave msg send --to claude --subject "Done" --message "Implemented X"

MCP server

agentweave mcp setup                                  # Configure in Claude + Kimi (once)
agentweave-mcp                                        # Run the MCP server (stdio)

Watchdog

agentweave-watch                                      # Monitor + print notifications
agentweave-watch --auto-ping --agent claude           # Auto-notify Claude on new messages
agentweave-watch --auto-ping --agent kimi             # Auto-notify Kimi on new messages
agentweave-watch --interval 3                         # Custom poll interval (seconds)

Transport (cross-machine)

agentweave transport setup --type git                 # Enable git transport
agentweave transport status                           # Show active transport
agentweave transport pull                             # Force immediate fetch
agentweave transport disable                          # Revert to local

Template Maintenance

agentweave update-template --agent claude --template-path ~/projects/template.txt
agentweave update-template --agent claude --focus "sub-agents"

---

Cross-Machine Collaboration

By default, AgentWeave works on a single machine. For cross-machine sync, enable git transport:

# Both developers run this (same git remote required)
agentweave transport setup --type git --cluster alice   # your workspace name

This creates an orphan branch (agentweave/collab) on your git remote. Messages and tasks sync through it using git plumbing — your working tree and current branch are never touched.

Messages are stamped alice.claude → bob.kimi so multiple workspaces can share the same remote.

The MCP server works transparently with all transports — it calls get_transport() internally.

---

Task Status Lifecycle

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

---

What Gets Created on Init

Project Root

AI_CONTEXT.md             # Project DNA — stack, architecture, standards (static)

.agentweave/ Directory

.agentweave/
├── AGENTS.md             # Collaboration protocol — MCP vs manual, workflow
├── ROLES.md              # Agent role assignments — who owns what domain
├── README.md             # Quick command reference
├── session.json          # Session config (gitignored)
├── shared/
│   └── context.md        # Current state — today's focus, recent decisions (dynamic)
├── tasks/
│   ├── active/           # JSON files for each active task (gitignored)
│   └── completed/        # Archived completed tasks (gitignored)
└── messages/
    ├── pending/          # Unread messages (gitignored)
    └── archive/          # Message history (gitignored)

File Purposes

File	Question It Answers	Change Frequency
AI_CONTEXT.md	What is this project? (stack, architecture, commands)	Monthly
AGENTS.md	How do we collaborate? (MCP vs manual, workflow)	Per session
ROLES.md	Who does what? (role assignments)	Per session
shared/context.md	What are we doing today? (focus, decisions, blockers)	Daily

---

Safety Features

File locking — prevents race conditions when both agents write simultaneously. Tasks and messages use file-based mutexes with a 5-minute automatic timeout.

Schema validation — all JSON state files are validated before saving. Agent names, task statuses, and required fields are enforced.

Input sanitization — string length limits and type coercion before any write.

Conflict-free git sync — GitTransport appends files with UUID-suffixed names; two machines can never produce the same filename.

---

Roles

Role	Example Agents	Responsibilities
Principal	claude	Architecture, planning, review, final decisions
Delegate 1	kimi	Backend implementation
Delegate 2	gemini	Frontend development
Delegate 3	codex	Testing, DevOps

In hierarchical mode (default), the principal assigns work and reviews results. In peer mode, agents can assign tasks to each other.

Roles are defined in .agentweave/ROLES.md — edit this file to customize responsibilities.

---

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.md and cluster naming
Local MCP server	Done (v0.4.0)	Native tool integration, zero-relay with watchdog pinger
AgentWeave Hub	Planned	Hosted MCP server, multi-team, web dashboard

The Hub (next phase) will be a hosted MCP server — any agent connects via HTTP, enabling real-time delivery without polling and a web dashboard for oversight. See ROADMAP.md for the full plan.

---

Installation Options

# From PyPI — base (manual relay mode)
pip install agentweave-ai

# From PyPI — with MCP server (zero-relay mode)
pip install "agentweave-ai[mcp]"

# From source
git clone https://github.com/gutohuida/AgentWeaveFramework.git
cd AgentWeaveFramework
pip install -e ".[mcp]"

---

FAQ

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. In MCP mode, even the relay step is automated.

Q: What's the difference between manual relay and MCP mode? In manual mode, you copy-paste a relay prompt from Claude to Kimi. In MCP mode, the watchdog fires a one-liner at the agent's CLI when a message arrives — no human involvement after the watchers are started.

Q: Do the watchdog processes need to stay running? Yes. Run agentweave-watch --auto-ping --agent <name> in a terminal (or as a background process / system service) for each agent you want auto-notified. If they're not running, messages still queue up — agents just won't be auto-triggered.

Q: How do delegate agents know how to use the system? agentweave init writes .agentweave/AGENTS.md — a complete guide covering commands, workflow, and protocol. In MCP mode, agents use the registered tools directly instead of CLI commands.

Q: Should I commit .agentweave/ to Git? Partially. The .gitignore excludes runtime state (tasks, messages, session.json, transport.json) but keeps AGENTS.md and README.md.

Q: Can I use this with a single agent? Yes — skip the relay step. Session, task, and summary commands are useful even for single-agent projects.

Q: Do both developers need the same git remote for cross-machine sync? Yes. Git transport requires a shared remote (e.g. origin). One developer creates the orphan branch with agentweave transport setup --type git, the other connects with the same command.

---

Links

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

---

MIT License