https://mcp.pome.sh/mcp. Transport is
Streamable HTTP only (no SSE resumability). It lets any MCP client author
tasks, launch evaluation runs, and read back scored reports, entirely through
tool calls — Connect to the MCP is where the clients and
their wiring live. This is what makes Pome usable by an AI agent directly, not
just a human at a CLI.
Coach and examinee
The agent you connect to the MCP is the coach. It authenticates, picks or authors a task, launches a run, and reads the report. The coach can see grading criteria. The examinee is a faithful clone of the agent-under-test that Pome launches separately — for example, on Anthropic Managed Agents — with its service MCP URLs swapped to Pome digital twins. The examinee holds zero Pome tools and never sees the criteria.run_task hands the coach a launch
spec for the clone; the coach launches it.
First call is always
get_started. It returns the coach playbook.Tool index
Every tool below links to its section in MCP tools, where each one is documented with the same inputs table, returns, errors, and example.Orientation
Authoring
Run lifecycle
Connect to the MCP
Add the server and authenticate.
MCP tools
Full reference for every tool.
Evaluate an agent over MCP
The end-to-end coach workflow.
Digital twins reference
The deterministic, resettable SaaS services
run_task provisions against.