tinyrouter.example.yaml

Rama main · réplica de lectura del repositorio original.

server:
  # Loopback by default: reachable from the network is worth deciding on
  # deliberately. Set it to 0.0.0.0 to serve other machines. The container image
  # sets TINYROUTER_HOST=0.0.0.0 itself, since a published port cannot reach a
  # process listening only on loopback inside it.
  host: ${TINYROUTER_HOST:-127.0.0.1}
  port: 8080
  api_key: ${TINYROUTER_API_KEY}
  # Or name several callers instead, so one can be revoked by deleting a line.
  # Set api_key or api_keys, never both.
  # api_keys:
  #   alice: ${ALICE_KEY}
  #   ci: ${CI_KEY}
  # /healthz, /readyz and /metrics are open by default, and disclose which
  # providers you use, how many tokens you spend, and which are failing right
  # now. Set this to put /readyz and /metrics behind api_key; /healthz stays
  # open so container probes keep working.
  protect_observability: false
  # Bun's own connection idle timeout, in seconds. 0 disables it. If set, it
  # must be longer than body_timeout_ms below, or it would close the connection
  # before that timeout could answer.
  idle_timeout_seconds: 0
  # Bounds the gap between request-body chunks: a slow upload is fine, one
  # that stops arriving is not.
  body_timeout_ms: 30000
  # Browser origins allowed to read the observability endpoints (/, /healthz,
  # /readyz, /metrics, /v1/models). Omitted means no response carries CORS
  # headers at all. Exact origins only - no "*", no paths - and completions are
  # never shared, so enabling this cannot let a page spend your provider credit.
  # allow_origin: ["http://localhost:5173"]

routing:
  retries: 0
  retry_statuses: [429, 500, 502, 503, 504, 529]
  # Retries of the same target wait for the provider's Retry-After when sent,
  # otherwise an exponential jittered delay; a single wait never exceeds
  # backoff_max_ms. Fallback to the next target is always immediate.
  backoff_initial_ms: 200
  backoff_max_ms: 2000
  # Demote a provider that keeps failing to a last resort instead of paying its
  # timeout on every request. Off by default; set failures to enable.
  circuit_breaker:
    failures: 0
    cooldown_ms: 30000

# Optional deterministic request filters, applied in order per attempted
# provider. A filter with a providers list applies only to those providers,
# so local targets can stay unfiltered. See "Request filters" in the README
# for the built-in pattern names.
# filters:
#   - type: block
#     patterns: [anthropic_api_key, openai_api_key, aws_access_key, private_key]
#   - type: redact
#     patterns: [email]
#     providers: [openai, anthropic, gemini]

providers:
  openai:
    type: openai
    api_key: ${OPENAI_API_KEY}
    # Bounds provider progress: this long for response headers, then this long
    # again for each gap between body chunks. A slow model is fine; a silent
    # provider is cut loose instead of hanging the client.
    timeout_ms: 60000

  anthropic:
    type: anthropic
    api_key: ${ANTHROPIC_API_KEY}
    timeout_ms: 60000
    # Claude Opus 4.7 and later reject a non-default temperature or top_p, and
    # the newest models also reject forced tool use, with an upstream 400 that
    # ends routing. If this provider's targets are such models, drop the one and
    # skip targets for the other. See "Anthropic provider options" in
    # docs/routing.md.
    # sampling_parameters: drop
    # forced_tool_choice: skip

  gemini:
    type: gemini
    api_key: ${GEMINI_API_KEY}
    timeout_ms: 60000

# Model IDs are passed to the provider verbatim and differ by account. If a
# target returns 404, list what your key can reach - see "Finding model IDs"
# in the README.
routes:
  fast:
    - openai/gpt-5-mini
    # Claude 4.5 models reject the adaptive thinking that reasoning_effort
    # becomes, and that 400 ends routing. If clients send reasoning_effort,
    # use a model that accepts it, such as claude-sonnet-4-6.
    - anthropic/claude-haiku-4-5

  smart:
    - anthropic/claude-sonnet-4-6
    - openai/gpt-5.2

  economical:
    - gemini/gemini-3.6-flash
    - openai/gpt-5-mini