Connect
Add the server URL in any MCP client that supports remote servers with OAuth:Write tools (creating datasets, starting training/evals) only work if you granted write at consent. To change it, reconnect and re-consent.
Tools
Explore
read
The queryable trace fields, their types/operators, and this project’s live intent labels + model names. Call it first.
read
Check a BQL filter for errors (with did-you-mean) without running it.
read
Run a BQL filter over the project’s traces — match count + a sample.
read
The project’s inefficiency + cost breakdown, aggregated in ClickHouse.
Datasets
read
List the project’s fine-tune (SFT) datasets with record counts.
read
A dataset’s training-data quality: intent distribution, redundancy, trajectory health.
write
Create a new empty SFT dataset.
write
Add
{messages, tools?} trajectories to a dataset.write
Build a dataset from real runs — every trace matching a BQL filter, reconstructed into a training record.
write
Export a dataset as fine-tune JSONL to bryel storage. Returns a
jobId.read
Poll an export job; once succeeded, a presigned download URL.
Training
read
The fine-tune presets (recipes): slug, base model, loop kind, expected data shape.
write
Start a fine-tune from a dataset + preset. Metered — pass
max_spend_usd to cap it. Returns a runId.read
Poll a training run: status, progress, loss, cost, and the weights download when done.
Evals
read
The eval suites (benchmarks) in the project: slug, name, default judge model.
write
Create one run per case × model for a suite. Returns each run’s
sessionId + case prompt for your harness to drive (see evals).read
A suite’s leaderboard over scored runs — overall, by model, and by case.
These are the same operations as the dashboard — the MCP server just lets an agent drive them.