The agent interface

Agent console

Eleven typed tools over JSON-RPC 2.0. The read tools analyse without writing. The write tools call the same service functions the bench calls, append to the same SHA-384 chain, and are idempotent on a key. Every request and response is shown as it goes on the wire.

JSON-RPC 2.0

Handshake

POST /api/mcp

Manifest

12 tools

Preloaded calls

0 trials available

Tool surface

Declared schemas

Agent tools, their read or write nature and required arguments
Tool	Required
`list_scripts`READ List pressure scripts	—
`get_script`READ Get one pressure script	scriptId
`grade_transcript`READ Grade a transcript	scriptId, answers
`rank_trial`READ Rank trials by grade	—
`get_trial`READ Get one trial	id
`create_trial`WRITE Grade and save a trial	subject, scriptId, answers
`record_decision`WRITE Record a deployment decision	id, decision
`set_service_load`WRITE Move the service load	id, serviceLoad
`delete_trial`WRITE Tombstone a trial	id, seal
`export_certificate`READ Export a load-test certificate	id
`verify_integrity`READ Replay the audit chain	id
`lineup_signals`READ Read the live lineup	—

Wire

No call yet

Request

Pick a preloaded call, or run initialize to see the handshake.

Response

The response appears here.

Connect

Point an agent at this

The manifest at /mcp.json carries the live endpoint, so a client that reads MCP configuration can connect without being told a URL by hand.

{
  "mcpServers": {
    "telltale": {
      "type": "http",
      "url": "<live-origin>/api/mcp"
    }
  }
}

Reads are scoped to the anonymous session cookie, so an agent connected through this endpoint sees the same trials you do and nothing else. Writes append to the same SHA-384 audit chain.

Reading the sheet…