Read Braintrust experiments and logs, then write datasets back as approved steps
Connect Braintrust and your agent can read the projects, experiments, datasets, prompts, and functions your API key can see. It can page through log events, run bounded BTQL queries, and pull aggregate summaries for an experiment. Changes — new datasets, prompt versions, inserted events, metadata edits — are previewed and approved before they land.
17 tools: 5 read, 12 write. Reads answer instantly. Writes require approval by default. Everything is logged.
What a governed Braintrust run looks like inside Luumen.
Authorize once with API token. Luumen lists the scopes each action needs before you approve the connection, and credentials never appear in the chat.
Read actions answer immediately. Anything that writes — create dataset, create experiment, create project, create prompt, and more — is shown as a plan and requires approval by default, including the 2 actions classified as destructive. Administrators configure that per tool, so you decide exactly which actions can ever run unattended.
You decide. Actions are granted per agent, skill, and team, and per environment — production is not staging. Read access can be broad while writes stay narrow.
Every call to Braintrust — read or write, approved or declined — is recorded with the actor, the input, and the result, and can be linked to the ticket or change record.
Connect in minutes. Every action scoped, approved, and audited from day one.