Percentile aggregations over traces are exactly the query a model gets subtly wrong — right metric, wrong grain, or a metric that does not exist in this project. Discovering the schema first turns that into a lookup. The platform tools deserve a second look before granting: platform_create_api_key means a connection with this tool can mint credentials, which is a different order of permission from reading traces.
LangWatch's endpoint over their LLM observability platform. Two servers exist: the main one covering traces, prompts and platform administration, and a separate governance server with its own nine tools.
- search_traces and get_trace — filters, metrics, aggregations and groups, over a date range or a relative window like 7d
- get_analytics — a named metric with an aggregation: average, sum, median, p90, p95, p99, cardinality or terms
- discover_schema — what is actually queryable, before writing a query
- list_prompts, get_prompt, create_prompt, update_prompt — prompt management
- Platform_* — projects: list, get, create, update, archive; and API keys: list and create
- fetch_langwatch_docs, fetch_scenario_docs, report_issue
OAuth authorization code with PKCE, scoped to your project. Free tier at zero forever; developer seats are €29 per month plus usage. Self-hosting is available at any time.
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen