Epinu MCP — agent door for a real-world asset marketplace

Epinu is an agent-first marketplace for real-world assets — agro and energy first, extending to mining, equipment, and industrial services. AI agents are first-class users: they can search, read, and draft deals directly over MCP, and every write goes through a human-approval broker with a full audit trail.

This is the public integration repository for Epinu's hosted MCP endpoint: registry metadata (ai.epinu/epinu in the official MCP Registry), runnable examples, and operational checks. The Epinu platform implementation is maintained separately.

• Endpoint: https://api.epinu.ai/api/agent/mcp (Streamable HTTP, JSON-RPC 2.0)
• Live docs for agents: epinu.ai/llms.txt
• Manifest: epinu.ai/.well-known/mcp

Try it in 60 seconds — no account, no token

bash examples/quickstart.sh

That queries the live public production endpoint anonymously and returns current public project and marketplace results. The sample queries currently surface an ASIC repair center in Odessa, TX and related services; inventory may change. (QA fixtures are excluded from public reads as of 2026-07-15.) The anonymous tier allows initialize, tools/list, and five read tools (getting_started, marketplace_listings_search, projects_search, search, and fetch) against public data — see TOOLS.md for the reference.

Or the same flow in plain Node (zero dependencies, Node ≥ 18):

node examples/walkthrough.mjs

Connect from Claude Code / any MCP client

claude mcp add --transport http epinu https://api.epinu.ai/api/agent/mcp

or in .mcp.json:

{
  "mcpServers": {
    "epinu": {
      "type": "http",
      "url": "https://api.epinu.ai/api/agent/mcp"
    }
  }
}

Works immediately in read-only mode. To let your agent draft deals, add "headers": {"Authorization": "Bearer epagt_..."} with a delegated token — see below for how to get one. Details in examples/claude-code.md.

The interesting part: writes never execute

Epinu's write tools (marketplace_listing_create, project_create, …) do not write to the database. They create proposals (status: pending) and return an approval_url. A human reviews the proposal in the Epinu dashboard and approves or rejects it — only approval executes the change.

examples/walkthrough.mjs demonstrates the agent-controlled portion of the flow: search → create a pending proposal → check its status. Human approval happens separately in the Epinu dashboard. examples/walkthrough.md is the recorded, redacted transcript of a real production run — proposal created, reviewed, and rejected by the human, with nothing ever written.

Why is every write human-approved?

• Agents make mistakes; markets are real. A hallucinated listing, a wrong price, or a duplicate is cheap to reject at proposal time and expensive to unwind after it's live. The broker makes the failure mode "a human clicked reject", not "a bad write hit production".
• Identity and audit. Every proposal records actor.type=personal_agent and the agent's client id. The human always knows which agent proposed what, and the trail survives.
• Scoped delegation, not account takeover. An epagt_ token is created by the account owner with explicit scopes (Dashboard → Settings → Authorize Agent). Out-of-scope calls fail with a clean scope error. Sensitive actions (deletes, profile changes) additionally require fresh MFA on the human side at approval time.
• **Drift

... View full README on GitHub