The failure mode for metered scoring is an agent iterating on payload shape and burning the month's allowance on rejected requests. Separating validation from scoring, and saying in the tool text that validation is free, removes it — the agent can converge on a correct payload and then spend once. get_workflow_schema serves the same end from the other direction, by supplying the field list rather than leaving it to be inferred. Reporting model versions on list_workflows is worth noting for anyone keeping scores comparable over time: a score without its model version is not comparable to last quarter's.
A workflow-scoring client with eight tools: the scoring calls themselves, the schema and validation around them, and the plan limits and usage reporting that govern them.
- score_workflow scores one item; score_batch handles up to 25 sequentially in a single request.
- validate_workflow_payload validates and normalises a payload and states that it does not consume a scoring call — so a malformed request costs nothing to discover.
- get_workflow_schema returns required fields, optional fields and example payloads for a workflow, which is what makes a first attempt correct.
- list_workflows returns the supported workflows with their current model versions, so a score can be attributed to a version rather than to 'the model'.
- get_limits and get_usage report the plan's ceiling and the consumption against it, with usage optionally scoped to a month.
- submit_feedback attaches structured feedback to a prior scoring request rather than to the model in general.
SENTINEL_BASE_URL and SENTINEL_TOKEN_BASE_URL — the base URLs are configurable rather than fixed, so this points at the deployment you are entitled to.
One command — uvx sentinel-signal-mcp