Buildkite converts job logs into a queryable form, so an agent can grep for the error with context lines or tail the last entries rather than pulling megabytes into the conversation. Everything needed to close the loop is in the same server — find the failing job, read the annotation, retry it — and token scopes stay the real boundary, with a read-only endpoint if you would rather nothing can be triggered at all.
Buildkite's own server over its REST API, covering pipelines, builds, jobs, logs, artifacts, annotations, clusters, agents and Test Engine.
- Builds listed for an organization or a pipeline as lightweight summaries, with a build's detail, its jobs filtered by state, and any single job by UUID
- Job logs made searchable rather than dumped: regex search with context lines, a tail of the last entries, and reads from a given row
- Recovery actions in the same place as the diagnosis — retry a job, unblock a blocked job, cancel or rebuild a build, or create a new one on a commit and branch
- Pipelines created and updated from YAML, and pipeline schedules read, created and updated with their cron expression, branch and enabled state
- Artifacts across a build or one job, with paths, sizes, MIME types and download URLs, and artifact content fetched base64-encoded
- Annotations listed for a build or job, and new ones created
- Test Engine data: tests in a suite with execution metrics over a time range, a single test's metadata for failed executions, and the runs for a suite
- Clusters and queues managed end to end, including pausing and resuming dispatch on a queue
- Agents inspected with their connection state, host, version, current job and pause status
- Responses trimmed to a subset of REST fields on purpose, to keep noise and token cost down
A Buildkite account. The recommended path is the remote server Buildkite hosts at https://mcp.buildkite.com/mcp, which issues a short-lived OAuth token for your user with read and write scopes; https://mcp.buildkite.com/mcp/readonly is the read-only equivalent, and https://mcp.buildkite.com/direct takes a Buildkite API token in an Authorization header for headless agents that cannot complete OAuth. To run it yourself, use the container image ghcr.io/buildkite/buildkite-mcp-server with BUILDKITE_API_TOKEN — each tool needs a matching token scope, such as read_builds or write_pipelines. Toolsets restrict which categories are exposed at all.