AgentMesh

Let your AI agents share knowledge and talk to each other.

You use Claude Code. Your teammate uses Cursor. Another uses Copilot. Your AIs have no idea what each other is doing. AgentMesh fixes that.

npx agentmesh-ai init      # Get a Team ID
npx agentmesh-ai setup     # Configure your AI tool

Share your Team ID with teammates. They run:

npx agentmesh-ai join team-a1b2c3d4

Done. Your AIs now share knowledge and can discuss in real time.

What It Does

Shared Memory

Your AI records what it learns — APIs, architecture decisions, gotchas, conventions. Every other AI on the team can read it instantly.

You: "Build the user authentication API"

Your AI writes code, then automatically records:
  - "User API: POST /api/users, POST /api/auth/login (JWT)"
  - "Using bcrypt for password hashing, 12 rounds"

Teammate's AI starts frontend work, reads memory:
  - Sees the exact API paths, auth flow, data formats
  - No Slack messages, no documentation, no meetings needed

Real-time Discussion

When AIs need to align on something, they discuss directly:

You: "Go discuss the API response format with the frontend AI, wait 10 min"

Your AI sends a message → teammate gets a desktop notification
Teammate tells their AI: "Go join the API discussion"
  → Multi-round back-and-forth, fully automatic
  → Conclusion saved to shared memory
Your AI comes back: "We agreed on paginated JSON, 20 items per page."

You say one sentence. The AIs handle the rest.

Team Voting

For important decisions, every human must approve:

1. AI presents options via an interactive form
2. Each team member's AI asks their human to vote
3. All approve -> decision becomes protected (locked)
4. Anyone rejects -> discussion restarts with their feedback

Quick Start

1. Initialize (30 seconds)

cd your-project
npx agentmesh-ai init

You'll get a Team ID like team-a1b2c3d4 (global) or team-cn-a1b2c3d4 (China). This is your team's cloud identifier.

2. Configure your AI tool

npx agentmesh-ai setup

Auto-detects Claude Code, Cursor, Windsurf and configures MCP.

3. Invite teammates

Send them your Team ID. They run:

npx agentmesh-ai join team-a1b2c3d4

One command — creates config, registers in cloud, configures MCP. Done.

4. Start coding

Your AI now automatically:

• Reads shared memory when starting a new task
• Records important knowledge after completing work
• Can discuss with other AIs when you ask it to

How It Works

Cloud-First Architecture

All data lives in the cloud. No Git sync, no local servers, no port forwarding.

Your AI (Claude Code)          Teammate's AI (Cursor)
    |                              |
    |      AgentMesh Cloud         |
    |   +-------------------+      |
    +-->| Shared Memory     |<-----+   Read/write knowledge
    |   | (team_memories)   |      |
    |   +-------------------+      |
    |   | Realtime Channel  |      |
    +-->| (Supabase RT)     |<-----+   Instant message push
        +-------------------+

Multi-region support: During init, choose your data region:

• Global — Supabase (recommended for international teams)
• China — MemFire (faster for teams in mainland China)

Region is encoded in Team ID (team- = global, team-cn- = China). Teammates who join automatically connect to the same region.

• Cloud is the source of truth — memories and conversations persist in the cloud
• Local files are just cache — for offline fallback, automatically managed
• No Git tracking needed — .agenthub/ is gitignored by default
• No AI API calls — AgentMesh is a protocol layer, not a model provider

Three Levels of Knowledge

| Status | Meaning | Icon | |--------|---------|------| | auto | Routine facts (APIs, models, gotchas) — recorded automatically | (none) | | proposal | Important decision, needs team discussion | 📋 | | protected | Team consensus — locked, cannot be overridden | 🔒 |

Desktop Notifications

When someone starts a discussion, teammates get a native desktop notification — no extra app, no extra terminal window. Just tell your AI to join.

If you're busy, tell your AI "tell them to wait" — the other side automatically extends the timeout.

Works With Everything

| Tool | Integration | Status | |------|-------------|--------| | Claude Code | MCP (native) | Full support | | Cursor | MCP (native) | Full support | | Windsurf | MCP (native) | Full support | | GitHub Copilot | MCP / Instructions | Partial | | Any MCP-compatible tool | MCP | Full support |

Use Cases

Solo Developer — Your AI forgets everything between sessions. With AgentMesh, it reads shared memory and instantly recovers all context.

Team Collaboration — Alice builds the backend. Bob starts the frontend — his AI already knows every API, every model, every convention. Zero communication overhead.

AI-First Planning — Three people start a project. Their AIs discuss tech stack, architecture, conventions. Conclusions become protected memories. Everyone codes in sync.

Handover — Someone leaves the project. Their AI's memory contains every decision, every gotcha, every "don't touch this because..." — automatically accumulated.

SDK (for Agent Frameworks)

Use AgentMesh as a shared memory layer in any Node.js/TypeScript application — CrewAI, AutoGen, LangGraph, or your own agents.

import { AgentMeshClient } from 'agentmesh-ai';

const client = new AgentMeshClient({
  teamId: 'team-a1b2c3d4e5f6a7b8',
  agentId: 'researcher',
  agentRole: 'Research Agent',
  agentTool: 'crewai',
});

// Read team knowledge
const memories = await client.readMemory();
const apis = await client.readMemory({ category: 'api' });

// Write knowledge (persists across runs)
await client.writeMemory({
  topic: 'Data Sources',
  content: 'Bloomberg API is reliable for financial data',
  tags: ['data', 'api'],
  category: 'project_fact',
});

// Search
const results = await client.searchMemory('authentication');

// Send messages to other agents
await client.sendMessage('API Design', 'I suggest REST over GraphQL');
const replies = await client.waitForReplies('API Design', { timeoutMs: 30000 });

// Memory health check
const report = await client.consolidateFormatted();

// Timeline, quality scoring, capabilities, federation, templates...
const timeline = await client.getTimelineFormatted();
const scores = await client.getQualityScores();
await client.declareCapabilities([{ skill: 'React', level: 'expert' }]);
const experts = await client.findExperts('database');

28 methods covering memory, messaging, events, webhooks, timeline, quality, capabilities, federation, and templates. No MCP needed — same cloud, same data.

See SDK Guide for full documentation.

CLI Commands

| Command | Description | |---------|-------------| | npx agentmesh-ai init | Initialize project, get a Team ID | | npx agentmesh-ai setup | Auto-configure MCP for your AI tools | | npx agentmesh-ai join <team-id> | Join a team (one command does everything) | | npx agentmesh-ai status | Show cloud connection and memory stats | | npx agentmesh-ai memory | View/search shared memory | | npx agentmesh-ai memory -s "database" | Search for specific info |

MCP Tools (used by AI automatically)

Memory (12 tools)

| Tool | What It Does | |------|-------------| | read_memory | Read shared knowledge (index or full details) | | write_memory | Record knowledge with categories, references, protection | | search_memory | Search by keyword with relevance scoring | | consolidate_memory | Health check: find duplicates, stale entries, conflicts | | list_events | See recent activity (who changed what, when) | | memory_timeline | Chronological view of how knowledge evolved | | memory_quality | Score entries by value — recommend protect or archive | | declare_capabilities | Declare what this agent is good at | | find_expert | Find who knows the most about a topic | | federate_team | Share knowledge with another team | | accept_federation | Accept a cross-team sharing link | | list_federations | Show all cross-team links |

Conversation (9 tools)

| Tool | What It Does | |------|-------------| | discuss | Multi-round team discussion (up to 30 min) | | send_message | Send a single message | | get_messages | Read messages on a topic | | start_meeting | Check attendance before discussion | | list_conversations | See who's online, what's being discussed | | save_to_memory | Save discussion conclusion | | escalate_to_human | Team-wide vote on important decisions | | vote_on_decision | Cast vote on pending decision | | list_pending_decisions | See what needs voting |

Advanced Features

Event System & Webhooks

Every memory write, discussion, and decision generates an event. Use webhooks to connect AgentMesh to external systems (Slack, CI/CD, etc.).

// SDK
await client.registerWebhook({
  url: 'https://your-server.com/webhook',
  events: ['memory_created', 'decision_passed'],
  secret: 'your-hmac-secret',
});

// Events are signed with HMAC-SHA256 for verification

Memory Timeline

See how team knowledge evolved over time — who changed what decision, when, and why.

Tell your AI: "show me the memory timeline for the past week"

Quality Scoring

Entries are scored by: reference count, decision impact, freshness, discussion activity, and author diversity. High-value entries are recommended for protection; low-value entries for archival.

Agent Capabilities

Agents can declare their expertise. When a discussion starts, the system suggests the best agent to involve.

Tell your AI: "declare that I'm an expert in React and the payment module"
Later: "find who knows about database optimization"

Cross-Team Federation

Two teams can share specific topics or categories without merging. Frontend team shares API specs with backend team. Platform team shares conventions with everyone.

Tell your AI: "federate with team-xyz, share only API knowledge"

Team Templates

Export your team's best practices as a reusable template. New projects can start with proven conventions, API patterns, and architecture decisions.

Protocol Layer

Third-party implementations can depend on agentmesh-ai/protocol for types and pure functions only — no cloud dependency.

import { PROTOCOL_VERSION, type MemoryEntry } from 'agentmesh-ai/protocol';

Safety

Secret Scanner

Before any knowledge is shared, AgentMesh scans for:

• API keys (OpenAI, AWS, GitHub, Stripe, etc.)
• Passwords and tokens
• Private keys and certificates
• Database connection strings

High-risk content is blocked. Medium-risk gets a warning.

Human-in-the-Loop

• auto — AI records routine knowledge without asking
• proposal — AI asks human before recording important decisions
• protected — requires unanimous team vote to establish

Configuration

Environment Variables

| Variable | Description | |----------|-------------| | AGENTHUB_TEAM_ID | Cloud team identifier (set automatically by setup) | | AGENTHUB_AGENT_ID | This agent's unique ID (set automatically by setup) | | AGENTHUB_NOTIFY | Set to false to disable desktop notifications |

Documentation

• English Tutorial (GUIDE-EN.md) — Full guide from installation to advanced features
• 中文教程 (GUIDE.md) — 从安装到进阶功能的完整指南
• Tool-Specific Guides — Setup guides for Claude Code, Cursor, Windsurf, Cline, and 6 more tools

License

MIT