generated: '2026-08-11'
method: derived
status: candidate
source: openapi/infer-by-flow7-public-api-openapi.yml
published_server: null
searched:
  note: 'Infer operates NO MCP server, and says so plainly rather than leaving it ambiguous. This file
    is a CANDIDATE tool shape derived from the published OpenAPI. It carries NO `type: MCPServer` pointer
    in apis.yml — wiring one would credit Infer with an agent surface it does not run.'
  probes:
  - url: https://infer.flow7.org/mcp
    method: POST tools/list
    status: 404
  - url: https://infer.flow7.org/.well-known/mcp.json
    method: GET
    status: 404
  - url: https://infer.flow7.org/.well-known/oauth-authorization-server
    method: GET
    status: 404
    note: Checked because an MCP server would normally advertise one; absent, consistent with no server.
  - url: https://infer.flow7.org/.well-known/oauth-protected-resource
    method: GET
    status: 404
  third_party: []
  llms_txt_tool_list: none
  note_on_llms_txt: /llms.txt lists resources and diagnostics, not MCP tools. It contains no tool inventory
    to fall back to.
agent_surface_actually_shipped:
  kind: Agent Skills
  discovery: https://infer.flow7.org/.well-known/agent-skills/index.json
  spec: https://schemas.agentskills.io/discovery/0.2.0/schema.json
  count: 3
  digests: SHA-256 per artifact
  note: This is the finding worth recording. Infer made a deliberate, documented choice to ship packaged
    Agent Skills — which run in the agent's own process against the caller's own key — instead of a hosted
    MCP server that would sit between the agent and the API and require Infer to hold or broker a credential.
    For a product whose entire pitch is spend accountability and never handling a user's key, that is
    a coherent architectural position, not a missing feature. See skills/.
candidate_tools:
- name: get_public_catalog
  maps_to: GET /api/public/catalog
  operation_id: getPublicCatalog
  auth_required: false
  description: Read published selectors, per-million prices, cache rates, privacy modes and per-model
    runtime capabilities. Optional `environment` query (sandbox | live).
  input_schema_source: parameters of getPublicCatalog
- name: get_public_status
  maps_to: GET /api/public/status
  operation_id: getPublicStatus
  auth_required: false
  description: Read overall and per-service operational state with p95 latency. Must be called before
    recommending or sending any paid request.
  input_schema_source: no parameters
- name: list_models
  maps_to: GET /v1/models
  operation_id: listModels
  auth_required: true
  scope: inference:write
  description: List selectors currently callable for the authenticated key's environment.
  input_schema_source: no parameters
- name: create_response
  maps_to: POST /v1/responses
  operation_id: createResponse
  auth_required: true
  scope: inference:write
  spends_money: true
  description: Create a Responses-style inference request against a canonical selector. Reserves the maximum
    estimated cost before routing and returns a receipt.
  input_schema_source: CreateResponseRequest + Idempotency-Key and X-Session-Affinity headers
  guardrails:
  - Would require explicit human confirmation — every successful call charges the wallet.
  - A failed call's Idempotency-Key is terminal; a retry needs a new key.
coverage:
  openapi_operations: 4
  candidate_tools: 4
  unmapped_operations: 0
deployment:
  mode: none
  verified: derived
  checked: '2026-08-12'
  source: catalog MCP census