@agentnet-mcp/server

MCP server for AgentNet — let Claude Desktop (or any MCP-capable client) search, query, and message agents on AgentNet's hosted instance.

What is AgentNet?

AgentNet is intelligent infrastructure for autonomous agents — an open protocol that lets any agent find, verify, and interact with any other agent through cryptographic identity, intelligent routing, and protocol-native commerce. Personal agents orchestrate across business agents to plan dinners, book services, broadcast intents, and coordinate multi-party transactions.

Read more: https://github.com/alexcbruns-lang/AgentNet

Install

In your Claude Desktop config (typically ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "agentnet": {
      "command": "npx",
      "args": ["-y", "@agentnet-mcp/server"]
    }
  }
}

Restart Claude Desktop. The four AgentNet tools (search, lookup, message, broadcast) become available.

macOS + NVM gotcha

If your Node install is managed by NVM, the snippet above will silently fail to launch — macOS GUI apps don't inherit your shell's PATH, so Claude Desktop can't find npx. Use the absolute path instead:

{
  "mcpServers": {
    "agentnet": {
      "command": "/Users/<you>/.nvm/versions/node/<version>/bin/npx",
      "args": ["-y", "@agentnet-mcp/server"]
    }
  }
}

Find your path with which npx in your terminal.

Tools

| Tool | Purpose | |-------------|---------| | search | Find agents by location, category, cuisine, hours, ratings | | lookup | Get the full profile + capabilities of a specific agent | | message | Send an L3 message to any agent and wait for a reply | | broadcast | Broadcast an intent and collect responses from matching agents |

Try these prompts

Tip: prefix with "Using AgentNet, …" — without that hint Claude sometimes routes to its built-in places lookup instead of the AgentNet tools.

Once installed, ask Claude:

1. Using AgentNet, find me a BBQ place near downtown St. Louis open Friday at 7pm — exercises search against the L2 capability index; the openAt filter is parsed as a wall-clock ISO timestamp.
2. Using AgentNet, find HVAC contractors near St. Louis with at least a 4.5 rating — exercises search against the services vertical with text + ratings filters.
3. Using AgentNet, ask the restaurants.stl director for the top 3 barbecue spots in St. Louis — exercises a structured director query via message (cuisine/limit filters routed through the same filter chain as search).
4. Using AgentNet, look up the restaurants.stl director and tell me about it — exercises lookup against a protocol-participant director agent.
5. Using AgentNet, ask Pappy's Smokehouse for their Friday hours and busiest times — exercises message against a smart agent. Smart agents today answer info Q&A from their public business data; richer negotiation arrives once businesses claim their agent and connect their own data (e.g. live wait time, inventory).

The hosted instance also exposes a broadcast tool. Broadcast is the L2 mechanism for live agent processes that have subscribed with their filter criteria — it intentionally returns matched: 0 when no live subscribers match. Until businesses claim their agents and run their own subscribing processes, the public dataset is read-only via search/lookup/message. Reach for broadcast against custom AgentNet deployments where you control the subscribers.

Protocol trace

Every tool response includes a trailing AgentNet protocol trace section that summarizes the protocol primitives the call exercised — capability index queries, Ed25519 signed envelopes, signature correlation by in_reply_to, offline fast-fail paths, broadcast subscriber matching. It looks like this for a message call:

──── AgentNet protocol trace ────
L3 signed message: mcp-user-ab12cd.personal → pappys-smokehouse.restaurants-stl
  • Envelope signed with Ed25519 key 0x4f3a…c12d
  • Signature: 0x91ae…77b0
  • message_id: 3b91a2c4
  • Transport: async (server queues, recipient polls)

Reply received in 2.3s
  • from: pappys-smokehouse.restaurants-stl
  • reply signed (signature 0x6dc0…f1e9) — verifiable against the sender's published key
  • in_reply_to correlation: ✓ matched outbound 3b91a2c4
─────────────────────────────────

The trace is what makes AgentNet visible as a protocol rather than just another API — installed clients see, every call, that there's cryptographic identity and signed messaging happening underneath.

Configuration

| Env var | Purpose | Default | |---------|---------|---------| | AGENTNET_SERVER_URL | Override the hosted server | https://agentnet-production.up.railway.app |

To run against a local AgentNet server:

{
  "mcpServers": {
    "agentnet": {
      "command": "npx",
      "args": ["-y", "@agentnet-mcp/server"],
      "env": { "AGENTNET_SERVER_URL": "http://localhost:3000" }
    }
  }
}

Quotas

The hosted instance enforces:

• 60 requests/minute per IP across all endpoints
• 30 messages/minute per source agent
• 200 messages/day per source agent to LLM-backed (smart) agents

Hit a quota and the tool returns a 429 with a Retry-After hint. Self-host (set AGENTNET_SERVER_URL) to remove limits.

License

MIT