Agent Hub

Free, open-source MCP server for real-time context sharing between AI coding tools.

Problem: Developer A (Claude Code) builds an API. Developer B (Cursor) needs those specs to build the frontend. Today: copy/paste through Slack. With Agent Hub: "Send my API endpoints to David's Cursor instance."

---

Transports

Agent Hub supports two transport modes:

Nostr (default) — agents communicate peer-to-peer via a public Nostr relay. No hub server required. Team membership is a shared secret; agent identity is a secp256k1 keypair generated on first use.

Claude Code ──[stdio]── mcp-client ──[WebSocket]──┐
                                                   ├── Nostr relay (e.g. wss://nos.lol)
Cursor      ──[stdio]── mcp-client ──[WebSocket]──┘

Hub — agents connect through a central Express server. Supports remote MCP (no local install needed for Claude Code) and provides a hosted public instance.

Claude Code ──[stdio]── mcp-client ──[HTTP+SSE]──┐
                                                  ├── hub-server
Cursor      ──[stdio]── mcp-client ──[HTTP+SSE]──┘

---

Setup

Option A — Nostr transport (default)

No hub server required. Add the client and call a setup tool:

claude mcp add agent-hub --scope user -- npx @agent-share/mcp-client

Then from within your AI session, call agent_hub_setup_create or agent_hub_setup_join. Your generated private key and team ID are saved to ~/.config/agent-hub/config.json — future sessions connect automatically.

The default relay is wss://nos.lol. Override with HUB_URL=wss://your.relay.

Known limitation: public Nostr relays may not persist custom event kinds. Messages sent to a disconnected agent can be lost if the relay drops them before the agent reconnects.

Option B — Hub transport

Use the hosted instance (no setup required):

The public hub is live at https://agent-hub-wild-glade-1248.fly.dev.

Or run your own locally:

npm install
NODE_ENV=development npm run dev

Or deploy to Fly.io:

fly volumes create agent_hub_data --size 1 --region iad --config packages/hub-server/fly.toml
fly deploy --config packages/hub-server/fly.toml

State persists to a Fly volume at /data. Verify after deploy:

fly ssh console --config packages/hub-server/fly.toml -C "cat /data/agent-hub.json"

2. Connect your agent

No credentials required at startup. Add the client, then run a setup tool from within the session:

claude mcp add agent-hub --scope user -- npx @agent-share/mcp-client

Then from within your AI session, call a setup tool to create or join a team (see Tools below).

To use the hub transport explicitly, set TRANSPORT=hub in your MCP client env or pass HUB_URL to point at a custom instance.

---

Tools

Available to your AI agent after connecting:

Tool	Description
agent_hub_setup_create(agentName)	Create a new team; returns an apiKey to share with collaborators
agent_hub_setup_join(apiKey, agentName)	Join an existing team using a shared key
agent_hub_send(to, type, content)	Send to a named agent or "broadcast" to all
agent_hub_list_agents()	List agents currently in your team
agent_hub_receive()	Fetch and clear your inbound message buffer
agent_hub_whoami()	Return your agent name, team ID, and connection info
agent_hub_reset()	Disconnect and erase stored credentials for this workspace; restart required to re-setup

Message types: api_spec, file_change, decision, todo, question

Example:

"Send my current API endpoint definitions to bob as an api_spec message."

Credentials are saved per-workspace to ~/.config/agent-hub/config.json. Each project directory gets its own team entry. Env vars (TEAM_API_KEY, AGENT_NAME, HUB_URL, TRANSPORT) take priority over the config file and are useful for CI/scripted use.

---

Configuration

All values live in config/*.props — nothing is hardcoded in source.

Section	Key	Default	Description
server	port	3000	Hub listen port
rate_limit	max_requests	100	Requests per minute
sse	max_message_buffer_size	100	Max buffered messages per agent
team	max_agents_per_team	20	Max agents per team
nostr	relay_url	wss://nos.lol	Default Nostr relay
nostr	heartbeat_ms	60000	Presence re-publish interval
nostr	presence_window_s	90	Subscription lookback window
hub	default_url	(hosted instance)	Default hub URL for the stdio client

---

Implementation

• packages/hub-server — Express + @modelcontextprotocol/sdk. Each SSE connection spawns its own McpServer instance so tools are scoped to the authenticated agent. Persistence uses a pluggable IStore interface: MemoryStore (dev), VolumeStore (Fly.io, active in prod), RedisStore (multi-machine). Swap by changing one import in store/index.ts.
• packages/mcp-client — Stdio MCP server. Transport is abstracted behind ITransport; HubClient (HTTP/SSE) and NostrClient (WebSocket + nostr-tools) are the two implementations. Starts in bootstrap mode if no credentials are found.
• Auth (hub) — API keys are 64-char hex, SHA-256 hashed at rest. Agents pass via ?api_key= or Authorization: Bearer.
• Auth (Nostr) — Permissionless. Team membership is a shared 32-char hex string used as a Nostr tag filter. Agent identity is a secp256k1 keypair generated on first use and saved to the local config.
• Tests — 36 unit tests for mcp-client (npm test -w packages/mcp-client), integration tests against a live relay (npm run test:integration -w packages/mcp-client).

---

License

MIT