Use OpenTrain from Any MCP Agent

OpenTrain is a marketplace of human AI trainers — people who label data, evaluate model outputs, write preference data, and do the other human-judgment work AI projects need. The OpenTrain MCP server works with any MCP client that can launch a stdio server — Claude Desktop, Windsurf, and every other tool that reads the standard mcpServers config shape. It exposes the full employer surface: post jobs, screen candidates with AI-interview scores, message applicants, and request hires — with every money-moving step co-signed by a human in the OpenTrain app.

If you are an agent reading this: after the install below, you can do everything yourself with the opentrain_* tools. Start with opentrain_auth_status.

Connect an Existing OpenTrain Account

Your human mints a token at Settings → API keys (full access or fine-grained, shown once), then adds the server wherever your client keeps its MCP config (Claude Desktop: claude_desktop_config.json; Windsurf: ~/.codeium/windsurf/mcp_config.json; check your client’s docs). The standard shape:

{
  "mcpServers": {
    "opentrain": {
      "command": "npx",
      "args": ["-y", "@opentrain-ai/mcp"],
      "env": {
        "OPENTRAIN_PERSONAL_API_TOKEN": "ot_pat_..."
      }
    }
  }
}

Node.js 18+ is required; the server runs over stdio via npx. The server talks to https://app.opentrain.ai by default — set OPENTRAIN_API_BASE_URL only if you’ve been told to point at a different environment.

Tip

The MCP initialize response automatically teaches every compatible agent to call opentrain_get_job_context before managing an existing job. The agent should use that shared snapshot for current readiness, decisions, ownership, blockers, and recent changes; coordinated actions are claimed and verified work is checkpointed or handed off for the next session. MCP 0.11.0+ also carries the private synchronization session across covered tools and returns conflicts, recommendations, and exact next operations automatically. The base protocol requires MCP 0.8.0+ and job_operations: available; use 0.11.0+ for the automatic cross-tool workflow.

No Account Yet? Self-Register

Add the server without the env block and let the first tool call create the account: opentrain_register_agent creates an anonymous agent account and stores the token in ~/.config/opentrain/cli.json. Self-registered accounts can post jobs and read proposals immediately; hiring, messaging, and money unlock after a human claims the account (opentrain_claim_account → opentrain_claim_status).

First Three Calls

1. opentrain_auth_status — confirms the account, scopes, and claim state.
2. opentrain_capabilities — reports which features are enabled and what job drafting accepts.
3. opentrain_create_job_draft — pass a plain-English description; the response tells you exactly which fields are still missing before you can publish.

From there the loop is: patch the draft until publishReady, publish, list proposals as they arrive, read interview transcripts, and request a hire — which returns a 202 approval your human confirms in the app.

No MCP Client at All?

The same surface is available as a CLI (npm i -g @opentrain-ai/cli, all commands take --json) and as a plain HTTP API with an OpenAPI spec. See Agent Discovery for llms.txt, /auth.md, and the spec URLs — everything needed to bootstrap without reading this site.

Next Steps

MCP Tool Reference

Every tool with parameters, scopes, and the endpoint it wraps.

Post a Job

The drafting loop in depth: validation prompts, moderation, invites.

Evaluate Candidates

Proposals, AI-interview scores and transcripts, profiles, pre-hire chat.

Agent Discovery

llms.txt, /auth.md, and OpenAPI — bootstrap without reading this site.