Braintrust MCP server

• Queries your logs, experiments and datasets with SQL

  Runs a SQL query over your Braintrust logs, experiments or datasets and returns the rows, or a download link when the result is large.

  Try: “Show me the last 10 logged requests with errors, with the error message and model used.”

  Read · sql_query

• Summarizes an experiment's scores and compares it to a baseline

  Pulls the aggregated scores and metrics for one experiment, optionally side by side with a baseline experiment so regressions stand out.

  Try: “Summarize my latest experiment and call out any score that regressed against the baseline.”

  Read · summarize_experiment

• Lists your recent projects, experiments, datasets and prompts

  Lists projects, experiments, datasets, prompts or functions that were created recently and that you can access.

  Try: “What experiments have I run recently in the chatbot project?”

  Read · list_recent_objects

• Breaks a long trace into its steps in time order

  Returns the meaningful model, tool and task steps of one trace in the order they happened, so the assistant can explain what a run did.

  Try: “Break down what this trace did, step by step.”

  Read · get_trace_work_items

• Finds recurring problems already reported in a project

  Searches a project's saved patterns by text, meaning, status or quality, including near-duplicates worded differently.

  Try: “What patterns are active in this project that I can fix right now?”

  Read · search_patterns

• Shows which fields your logs or datasets have and common values

  Discovers the fields, data types and most common values in your experiments, datasets or logs before a query is written.

  Try: “What fields are in my production logs, and which are filled in most often?”

  Read · infer_schema

• Lists a project's saved dashboards

  Lists the saved monitor dashboards in a project and the charts on each.

  Try: “Which dashboards does this project have?”

  Read · list_monitoring_views

• Shows what a saved dashboard contains

  Reads one saved dashboard, including its settings and every chart definition in order.

  Try: “What charts are on my error monitoring dashboard?”

  Read · get_monitoring_view

• Previews a chart on your logs without saving it

  Draws a preview chart over project logs, such as daily cost or p95 latency, without changing any saved dashboard.

  Try: “Chart daily cost for this project over the last two weeks, but don't save it yet.”

  Read · generate_monitor_chart

• Lists a project's alerts, scoring rules, exports and other automations

  Lists every automation in a project, such as online scoring rules, alerts, exports, retention policies and Topics, with full settings.

  Try: “What automations are set up in this project, and which ones are paused?”

  Read · list_automations

• Lists saved review views for a project's datasets

  Lists up to 100 saved custom views for labelling dataset rows, with each view's code and a link to open it.

  Try: “Show me the custom review views saved on my datasets.”

  Read · list_dataset_custom_views

• Shows a project's settings and default preprocessor

  Reads a project's settings, including which preprocessor facets and other functions use by default.

  Try: “What preprocessor is my project using by default?”

  Read · get_project_settings

• Tries a scorer on real traces without saving results

  Runs a saved, built-in or draft scorer against chosen spans or traces and shows the scores, without writing anything back.

  Try: “Test this helpfulness scorer on ten recent traces.”

  Runs code · test_evaluator

• Tries a preprocessor on real traces without changing them

  Runs a preprocessor on up to 50 spans or traces to preview the text it produces, without writing to the traces.

  Try: “Test the preprocessor you wrote on a few of my traces.”

  Runs code · test_preprocessor_on_trace

• Tries a facet on up to ten traces without saving the result

  Runs a draft facet definition on up to ten spans or traces to preview what it extracts, without writing to them.

  Try: “Try this facet on a few traces and show me what it pulls out.”

  Runs code · test_facet_on_trace

• Makes a shareable link to a Braintrust object

  Creates a direct web link to an experiment, log entry, dataset, prompt or other object so you can share it.

  Try: “Give me a link to those experiment results so I can post it for my team.”

  Read · generate_permalink

• Finds teammates by name or email and matches IDs to people

  Turns organization member IDs into names and emails, or searches members by partial name or email.

  Try: “Who logged the traces in this project this week?”

  Read · lookup_users

• Matches API key IDs to their names

  Turns API key and service token IDs into their names, so you can see which key produced which traces.

  Try: “Which API keys produced the traces in this project?”

  Read · lookup_api_keys

• Gets short-lived download links for files attached to logs

  Exchanges up to 100 attachment references from query results for download links that expire after five minutes.

  Try: “Get me download links for the images attached to these log rows.”

  Read · resolve_attachments

• Runs an eval comparing prompts or tasks on a dataset

  Runs an experiment with a dataset, inline rows or a past experiment as input, any task and any scorers. It can run your code and call AI providers, which uses compute and model credits.

  Try: “Run an eval comparing these two prompts on my regression dataset and tell me which wins.”

  Runs code · run_eval

• Copies log rows into a dataset

  Copies chosen project log rows into a dataset in the same project, keeping their input, metadata, tags and a link back to the log.

  Try: “Add the traces that failed the helpfulness scorer to my regression dataset.”

  Write · add_to_dataset

• Adds, edits or deletes rows in a dataset

  Inserts, updates or permanently deletes dataset rows, and can create a new dataset by name. Review deletes before approving them.

  Try: “Set the expected output on these dataset rows to the corrected answers.”

  Delete · edit_dataset_rows

• Saves a prompt as a new or updated version

  Creates a versioned prompt from a completion-style prompt or chat messages, or saves a new version of an existing one.

  Try: “Save this prompt to my project as a new version of the summarizer prompt.”

  Write · create_prompt

• Saves a new scorer or LLM judge

  Creates a versioned scorer, either an LLM judge or inline code, that returns a score or a category label.

  Try: “Create an LLM-as-a-judge scorer for helpfulness.”

  Write · create_evaluator

• Sets up a rule that scores production logs automatically

  Saves or replays an online scoring rule that runs saved scorers on production logs. New rules start paused.

  Try: “Set up online scoring with the scorer you just created, and leave it paused.”

  Write · update_online_scoring_rule

• Records a recurring problem with the evidence behind it

  Saves a new pattern, a recurring behavior found in traces, with supporting evidence and first and last seen times.

  Try: “I think we're overspending on web searches. Dig in and report a pattern if you find one.”

  Write · new_pattern

• Updates a reported pattern or marks it as a duplicate

  Edits a saved pattern's severity, category, quality or feedback, attaches more trace evidence, or closes it as a duplicate.

  Try: “Mark this pattern as a duplicate of the one we already have.”

  Write · update_pattern

• Saves step-by-step notes back onto a trace

  Saves the work sections and per-step annotations for a trace, so the breakdown stays with it.

  Try: “Save your step-by-step breakdown on this trace.”

  Write · update_trace_work_report

• Creates a new dashboard

  Creates a project dashboard, optionally with charts you already previewed.

  Try: “Create a cost dashboard with total cost, cost per trace and cost by model.”

  Write · create_monitoring_view

• Adds, changes or removes charts on a dashboard

  Inserts, edits or removes charts on an existing dashboard, one at a time or several at once.

  Try: “Add a p95 latency chart to my error monitoring dashboard.”

  Delete · update_monitoring_view

• Alerts you when a log matches a condition

  Creates an alert that fires on individual matching project logs, with an optional limit on how often it notifies.

  Try: “Alert me whenever a log has a refusal from the support bot.”

  Write · create_log_alert

• Alerts you when a metric crosses a threshold

  Creates an alert on an average, count, rate or percentile over a recent window, checked on a schedule.

  Try: “Alert me when the error rate goes above 2% over the last hour.”

  Write · create_threshold_alert

• Alerts you when an environment is updated

  Creates an alert for environment updates, optionally limited to certain environments.

  Try: “Alert me whenever the production environment changes.”

  Write · create_environment_update_alert

• Schedules a recurring analysis of recent traffic

  Creates a Loop job that analyzes a recent window of project data on an interval or cron schedule.

  Try: “Every Monday morning, analyze last week's traces and summarize the top failure modes.”

  Write · create_scheduled_loop_job

• Pauses or turns on an alert, job or scoring rule

  Pauses or resumes an alert, scheduled job or online scoring rule.

  Try: “Pause the online scoring rule you just created.”

  Write · set_automation_status

• Adds a new score type for human reviewers

  Creates a category, slider or free-text score that human reviewers fill in for a project.

  Try: “Add a 1 to 5 accuracy slider for human reviewers.”

  Write · create_human_review_score

• Sends traces or dataset rows to human review

  Adds up to 100 log traces, experiment traces or dataset rows to the human review queue, optionally assigning reviewers.

  Try: “Send these 20 failing traces to human review and assign them to Dana.”

  Write · add_to_human_review_queue

• Saves a custom view for labelling a dataset

  Saves a custom React view on a dataset for reviewers to label rows, and returns a link that opens it.

  Try: “Save a review view on this dataset that shows the question and answer side by side.”

  Write · save_dataset_custom_view

• Saves code that turns raw traces into readable text

  Creates a versioned preprocessor from JavaScript that converts raw trace data into text for Topics and facets.

  Try: “Write a preprocessor that works with my trace shape.”

  Write · create_preprocessor

• Saves a facet that pulls a short summary from each trace

  Creates a versioned facet that extracts a short summary from spans or traces using Braintrust's facet model.

  Try: “Create a facet that extracts what the user was trying to do.”

  Write · create_facet

• Turns on Topics for a project

  Enables Topics for a project so new traffic is grouped into topics. It does not process past traces.

  Try: “Set up Topics for my project.”

  Write · enable_topics_automation

• Changes a project's Topics settings

  Updates a Topics automation's facets, topic maps, filters, sampling or timing, and can remove facets and their topic maps.

  Try: “Add a sentiment facet to my Topics setup.”

  Delete · set_topics_automation

• Reprocesses past traces for Topics

  Runs a Topics automation over historical traces from a start time or recent window. This draws from your monthly model credits.

  Try: “Rewind my Topics automation to cover the last 30 days.”

  Runs code · rewind_topics_automation

• Lets Topics use the latest version of each topic map

  Removes version pins from a Topics automation's topic maps so new traffic uses the latest version of each.

  Try: “Unpin my topic maps so new traffic uses the latest versions.”

  Write · unpin_topic_maps

• Changes a project's default preprocessor

  Sets or clears the preprocessor that facets and other project functions use by default.

  Try: “Make the preprocessor you just created the project default.”

  Write · set_project_default_preprocessor

• Searches Braintrust's documentation

  Searches Braintrust's docs for guides, API references and code examples so answers come from the docs.

  Try: “How do I create a custom scorer in Braintrust?”

  Network · search_docs

• Sends a problem report about the server to Braintrust

  Reports a problem with the MCP server, such as a tool error or missing feature, to the Braintrust team. It changes no data.

  Try: “Tell Braintrust that the dashboard tool keeps failing on my project.”

  Network · send_feedback

• Used by the assistant behind the scenes

  Looks up connected Slack channels for alert delivery

  Lists the Slack workspaces and public channels connected to a project, so alerts can be sent to the right channel.

  Read · list_slack_channels

• Converts names to IDs and reads Braintrust links

  Translates between object names and IDs and works out what a Braintrust URL points to. The assistant uses it before other tools.

  Read · resolve_object

• Loads a step-by-step guide for a Braintrust workflow

  Loads a Braintrust workflow guide that tells the assistant what order to do things in and what to check at each step.

  Read · load_braintrust_skill

Watch out for

• It can change and delete your Braintrust data

  The assistant can create prompts, scorers, alerts and datasets, and can permanently delete dataset rows. Set your client to ask before each write action, and review deletes before you approve them. docs.braintrust.dev

• Running evals uses compute and model credits

  Evals can run your code and call AI providers, and reprocessing old traces for Topics draws from your monthly model credits. Ask for a preview before large runs. docs.braintrust.dev

• An API key acts as its owner

  With an API key, every action runs with the key owner's permissions. Prefer signing in with your own account so each person acts as themselves. docs.braintrust.dev

Security advisories