Skip to main content

Motivation

Evals have to run somewhere. A real suite means hundreds of sandboxes starting at once, agents acting inside them for minutes at a time, and a graded record of every attempt that still makes sense a month later. The platform is the hosted service that does this. You point the SDK or the hud CLI at it, it places each attempt on its own infrastructure, and it keeps the result as a trace you can reopen, share, and compare.

One API behind everything

The SDK, the CLI, and hud.ai are all clients of the same HTTP API, and nothing is reserved for the UI. Anything the platform can show you, your own code can fetch: a dashboard of your team’s scores, a bot that posts last night’s results, a CI check that fails a build when a score drops. The platform works as a backend you build on, not just a site you visit. Base URL https://api.beta.hud.ai, every public route under /v2, authenticated with a project HUD_API_KEY.

Reading the docs

  • REST API - the whole HTTP surface, grouped by what each part is for, with a request and response example for every endpoint.
  • MCP - a read-only Model Context Protocol server so coding agents can query environments, tasksets, jobs, and traces with the same API key.
For exact schemas and field-level detail, the generated reference is always in sync with the server. For an agent-readable version, point your agent at https://api.beta.hud.ai/openapi.json, the OpenAPI document both are built from. To attach the platform as tools in an agent host, use the Platform MCP at https://api.beta.hud.ai/v2/mcp/.