Braintrust MCP Server

Braintrust is an evaluation and observability platform for LLM applications: you run experiments, score them, and ship logs from production into the same place. Its MCP server closes the loop by letting the agent that is writing the code also interrogate the results, without a context switch into the web UI.

The centre of it is sql_query, which runs SQL directly against experiments, datasets and logs, with infer_schema to discover the available fields first so the agent is not guessing at column names. summarize_experiment returns aggregate scores and can diff them against a baseline experiment, which is the question you actually want answered after a prompt change — did this get better, and by how much.

Around that sit the navigation tools: resolve_object converts between names and IDs and parses Braintrust URLs, list_recent_objects surfaces recently touched projects, experiments, datasets, prompts and functions, and generate_permalink produces a shareable link so a finding from an editor session can go straight into a pull request or a Slack thread. search_docs covers Braintrust's own guides and API reference, and search_patterns, new_pattern and update_pattern record recurring project patterns.

It is a hosted remote server at https://api.braintrust.dev/mcp, with an EU data plane at https://api-eu.braintrust.dev/mcp. Authenticate either with an API key as a bearer token or interactively through OAuth 2.0, no key configuration required.

The workflow it enables: debugging a production log without leaving the IDE, comparing two experiment runs, checking cost metrics per model, and sharing the evidence with a link rather than a screenshot.