Skip to main content
Ask your own questions of your results, such as pass counts by treatment, cost by agent, or Evals attempts by model. You send a JSON query (never SQL) over a fixed set of datasets, and Obss returns bounded rows. For ready-made summaries, start with obss results analyze. Use data query for what it doesn’t answer, or when its page.completeData points here.

obss data describe

List the datasets you can query, with their fields, the filters and aggregates each field takes, default order, and limits.
Plugin (MCP) tool: inspect_data_model Takes only the global flags. Examples Read the model before you write a query:
Result fields Notes The model is the same in every project.

obss data query

Run a JSON query over one dataset.
Plugin (MCP) tool: query_data Flags Pass one of the two. Also takes the global flags. Examples Count trials by outcome:
Compare passed trials by treatment in one experiment:
Count Evals attempts by mode and Evals model:
Average cost per trial by agent, split by cost source:
Result fields Notes While hasMore is true, send the query again with nextCursor as its cursor. When truncated is true, reasons says why the page holds less than everything that matched: A query that doesn’t match the model fails with a usage error that names each problem, such as an unknown field or an operator the field doesn’t take.

Query JSON

Aggregate functions are count, count_distinct, sum, avg, min, and max; obss data describe lists what each field takes. Timestamps are ISO 8601 strings. A query runs for at most 3 seconds and returns at most 200 rows and 20 KiB.

Datasets

Fields to know

The frictions fields category, phase, confidence, severity, classification, and owner are always null and stay only so older queries run. No dataset holds message text or tool input and output; read those with obss runs trajectory.

Next steps