About Agentishare
Agentishare is an MCP server published by JTP75 in the Developer Tools category: uh um uh yeah.
The repository has 1 stars and 0 forks, with the last commit 7 months ago. Six months or more without a commit doesn't mean the server is broken, but check the open issues before depending on it in production.
Install
git clone https://github.com/JTP75/agentishareThis server has no ConduID identity, so agent calls to it are not receipted. Pin the version you install and review the source before granting it credentials.
Tools (7)
agent_hub_list_agents()List agents currently in your team
Parameters not published.
agent_hub_receive()Fetch and clear your inbound message buffer
Parameters not published.
agent_hub_reset()Disconnect and erase stored credentials for this workspace; restart required to re-setup
Parameters not published.
agent_hub_send(to, type, content)Send to a named agent or `"broadcast"` to all
Parameters not published.
agent_hub_setup_create(agentName)Create a new team; returns an `apiKey` to share with collaborators
Parameters not published.
agent_hub_setup_join(apiKey, agentName)Join an existing team using a shared key
Parameters not published.
agent_hub_whoami()Return your agent name, team ID, and connection info
Parameters not published.
Ask AI
Ask AI about Agentishare
Powered by Claude · Grounded in docs
Security checks
Static checks derived from the repository 5 days ago. Heuristics on the source, not an audit.
- ✓README presentThe repository ships documentation.
- ✓License declaredUsable under a stated license.
- ✓Tests presentA test suite exists in the repository.
- ✓Dependencies pinnedA lockfile pins dependency versions.
- ✓No dynamic code executionSource scan found no eval, dynamic Function, or shell execution.
- !Scoped permissionsDoesn't declare a permission scope. Assume it can do anything its process can.
README
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 ownMcpServerinstance so tools are scoped to the authenticated agent. Persistence uses a pluggableIStoreinterface:MemoryStore(dev),VolumeStore(Fly.io, active in prod),RedisStore(multi-machine). Swap by changing one import instore/index.ts.packages/mcp-client— Stdio MCP server. Transport is abstracted behindITransport;HubClient(HTTP/SSE) andNostrClient(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=orAuthorization: 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
README fetched from the repository 5 days ago. The original is authoritative.