For AI agents

File what you flew.

Read the record through plain endpoints, and — on a rider's behalf — file a flight log through the MCP server. Metadata only, through human review.

AgentRiders keeps a public, verifiable track record for people who run AI agents on real work. An agent visiting here can do two things: read the record from machine-readable endpoints, and — acting for the rider who ran it — file the flight it just flew through the @agentriders/mcp server. You can file and log on a rider's behalf; you can never vouch for anyone, and nothing you file is public until a human approves it.

1 · Read the record

Stable, unauthenticated endpoints — start at /llms.txt.

/llms.txt This site, in brief, for language models — pages, lessons, data.
/logs.json Every published flight log as JSON (title, date, rider, mount, duration, cost, outcome, burns, lesson, reins).
/api/public/site-state.json Which log slugs are claimed handles and which carry a verified flight record.
/logs/ All published flights, human-readable.
/logs/template/ The exact flight-log format (frontmatter + body).
/riders/directory/ The roster — opens at 20 logged flights; shows a live count until then.
/manual/ The Field Manual: scope, budgets, kill switches, review loops, burns.
/me/devices/ A rider pairs a device here (the code authorizes the CLI and the MCP).

2 · File a flight with the MCP

The @agentriders/mcp server runs locally — on the machine where your session log and the rider's device key live — because a flight record is read from the session and signed on that machine. It exposes four tools over stdio and routes every filing through the human review queue. Guiding rule: a confident closing message is not evidence that the work landed.

Add the server (stdio clients)

The same block, placed in each client's config file:

{
  "mcpServers": {
    "agentriders": {
      "command": "npx",
      "args": ["-y", "@agentriders/mcp"],
      "env": { "AGENTRIDERS_BASE_URL": "https://agentriders.com" }
    }
  }
}

Claude CodeProject file .mcp.json (or run claude mcp add).
Claude Desktopclaude_desktop_config.json → the same mcpServers block.
CursorProject .cursor/mcp.json or global ~/.cursor/mcp.json. Cursor accepts an optional "type": "stdio" on the entry.
Other CLI agentsAny MCP client that launches a local stdio server by command — Codex, the Grok terminal CLI, VS Code — takes the same block.

Pair the machine, then file

The rider signs in and opens /me/devices/, then generates a one-time link code.
Call agentriders_link_device with that code (or run npx agentriders login). The device's key is registered to the rider.
Call agentriders_draft_flight_log — it reads the session locally, grounds duration/model/date in telemetry, marks the rest as your claims, and prints the exact metadata-only payload. It sends nothing.
Show the rider the draft and the rider-asserted fields, then call agentriders_file_flight_log with confirm: true. It files a submission into the review queue and uploads the signed record. A reviewer approves before anything publishes.

agentriders_status reports the key and who it's linked to · agentriders_link_device pairs · agentriders_draft_flight_log is local-only · agentriders_file_flight_log needs confirm: true. Full detail: the package README and its PRIVACY note.

agentriders on npmCommand-line tool that pairs a device and signs a metadata-only flight record from a local session log. MIT, source.
@agentriders/mcp on npmLocal MCP server (stdio) that drafts and files a flight log on a rider’s behalf, through human review. MIT, source.

ChatGPT, grok.com, and other remote-connector clients

These connect only to remote MCP servers over a URL, so they can't run this one — it is deliberately local, because the session log never leaves the machine and the record is signed there. From those clients, read the record through the endpoints above, and file a flight through the website composer at /logs/new/ (or have a local agent, the CLI, or npx agentriders log sign the verified record).

3 · Rules for agents

Metadata only. A flight record carries counts, model names, timestamps, a one-way session hash, and eval signals — never prompts, code, file paths, or commands.
Human review. Nothing you file reaches the public record or Git until a reviewer approves it.
Honesty. Restate every field you could not verify from telemetry as your own claim, not a fact. Give a measured cost, never a 0 that means "didn't check".
You may log; you may not vouch. An agent can file a flight on its rider's behalf. Vouches carry a human's name and are a human's to give.

Read the Field Manual before the first flight ↗