Wire kestrel into your agent (MCP)

What MCP wiring buys you

The Model Context Protocol lets an agent use a tool without bespoke integration. kestrel speaks it, so once you wire the server into your client the agent drives the platform natively — the same operations the canonical HTTP face exposes, projected as MCP tools. This is a practicum: read the four faces for why MCP is an equal projection, and the MCP install page for the reference surface. Here we connect it.

Connect in one command

The kestrel MCP server is a hosted Streamable-HTTP endpoint — https://mcp.kestrel.markets/mcp. There is nothing to install and no key to paste. In Claude Code or Claude Desktop:

claude mcp add --transport http kestrel-markets https://mcp.kestrel.markets/mcp

Confirm it connected:

claude mcp list

For any client configured by file (Claude Desktop's claude_desktop_config.json and other Streamable-HTTP-aware clients such as Cursor), add the server directly:

{
  "mcpServers": {
    "kestrel-markets": {
      "type": "http",
      "url": "https://mcp.kestrel.markets/mcp"
    }
  }
}

No headers are required; the bare command connects on the first try.

What the agent gets

After initialize, tools/list returns the full projected tool set. The seven free-path operations are the subset you reach anonymously — each a thin equal projection of its HTTP endpoint:

capability_discovery — the always-available self-description of the platform
catalog — browse the free derived-data catalog
validate — check a Kestrel document without spending (the document travels as source)
simulate — run a metered, certified sim over a catalog dataset
grade — the signed black-box receipt every sim certifies
proof — resolve a public, re-verifiable certified proof URL
offers — the continuation quoted against work already underway

The list also carries the capability-gated Firm surface — pod, fleet, and agent tools that are never anonymous and out of scope for the trial funnel — so the live tools/list is larger than these seven. The seven above are the free-path subset.

The agent is admitted with an anonymous trial by default: no key, no signup. It can discover the tools, mint a trial, run a free sim, and reach a certified proof URL — the whole free path — over MCP alone.

Registering the server is the residency ask

Wiring the server once is what makes the capability resident. The registration persists in your client config, so the tools are there in the next session and the one after — no fresh discovery hop, no re-adding the endpoint. That persistence is the point: the agent keeps a standing door to the platform rather than rediscovering it each time.

Where OAuth enters

The anonymous trial can never carry paper, broker, live, or wallet scope. Settling an Offer or resuming an Operation under a paid or irreversible scope authorizes over OAuth 2.1 — a wallet signs commerce-only scopes, a human signs the identity-bound, legally irreversible ones. The free simulate-and-prove path above needs none of it.

A worked example

An agent with the server wired runs a free sim on a generic index scenario and mints a proof URL — entirely over MCP, no account. It hands the link to its human, who re-verifies the signature in a browser. Next session, the tools are still there: the residency ask means the agent never re-discovers the platform.

See it in kestrel

Confirm the endpoint is live and the receipt is real from the CLI, then wire the server:

npx kestrel.markets verify https://kestrel.markets/proof/art_66d7dda7f0466f69c123463c

That re-verifies a certified proof against independently fetched keys — the same proof the wired agent can mint for itself. Add the MCP server with the command above and the capability returns every session, no account in between.