This profile is part of our MCP server directory, within The MCP Guide. Profiles are checked against each vendor's own documentation. Need a server built? See MCP development.
The Buildkite MCP server lets an AI tool read pipelines, builds, jobs, logs and test results, and take actions such as retrying failed jobs. Buildkite publishes it as a hosted remote server with OAuth and as an open-source local server.
Checked against: Buildkite MCP server docs on 28 October 2026. eSEOspace did not install or test this server.
Key Takeaways
- The hosted server runs at
https://mcp.buildkite.com/mcpwith OAuth, andhttps://mcp.buildkite.com/mcp/readonlyis the read-only version. - A local server is available as a Docker image (
buildkite/mcp-server), a release binary or from source. - Buildkite says the server "is under active development," so tools may change, including breaking changes.
- OAuth access tokens last up to one hour and refresh tokens up to 30 days.
- Buildkite warns that retries and build starts can run pipeline commands, including deployments.
At a glance
| Item | Detail |
|---|---|
| Publisher | Buildkite |
| Official? | Yes |
| Hosting | Both: remote URLs and local Docker image or binary |
| Transport | HTTP (remote), stdio (local) |
| Sign-in | OAuth (remote), API access token (local or /direct endpoint) |
| Read or write | Read and write; read-only endpoint available |
| Price | No separate MCP price listed |
| Source code | buildkite/buildkite-mcp-server (MIT) |
| Status | Under active development (no beta or GA label) |
What it can do
Buildkite organizes tools into toolsets. The docs and README name toolsets such as investigations, builds, logs, annotations, tests and retries.
Tools named in the GitHub README include:
- Builds:
list_builds,get_build,wait_for_build,list_jobs,get_job - Logs:
search_logs,tail_logs,read_logs - Investigations:
compare_builds, which answers "what changed since this build last worked on main?" - Retries:
retry_job,retry_failed_jobs(with adry_runoption) - Pipelines:
create_pipeline,update_pipeline,get_pipeline,list_pipelines - Tests:
list_test_workflows,get_test_workflow
The retries toolset is built for least privilege. It lets an agent investigate and retry failures with the read_builds, read_build_logs and retry_builds scopes, without broader write_builds access.
How to connect it
Buildkite's remote setup page gives this Claude Code command:
claude mcp add --transport http buildkite https://mcp.buildkite.com/mcpFor a read-only connection limited to certain toolsets:
claude mcp add --transport http buildkite-read-only-toolsets https://mcp.buildkite.com/mcp/readonly --header "X-Buildkite-Toolsets: user,pipelines,builds"You can pin an organization with ?organization=acme on the URL. Headless agents that cannot complete OAuth can use https://mcp.buildkite.com/direct with an Authorization: Bearer <buildkite-api-token> header.
For a local server, the README shows starting a single toolset:
buildkite-mcp-server stdio --enabled-toolsets retriesBuildkite suggests pinning a local version for stable behavior, because hosted endpoints update automatically.
Security notes
- Read-only endpoint.
/mcp/readonlyexposes only tools whose required scope starts withread_, plus tools with no scope. - Not a sandbox. Buildkite says "Read-only MCP access restricts this connection, not the AI agent's other tools or credentials." An agent with shell access and a separately signed-in
bkCLI can still make changes. - Writes can deploy. Retries and build starts can run pipeline commands, including deployments. Buildkite advises requiring approval for writes.
- Local tokens. The local server uses a plain-text API token that you must protect.
- Rate limits. The
/directendpoint allows 50 requests per minute per user. Local server requests count toward your organization's REST API limit.
The MCP project's security best practices also cover risks such as token passthrough, confused deputy attacks and scope minimization.
See excessive permissions and approval fatigue, and the MCP Security Center.
Alternatives
- CircleCI MCP server for teams on CircleCI.
- GitHub MCP Server, whose
actionstoolset covers GitHub Actions runs and job logs. - GitLab MCP server, whose
citoolset covers GitLab pipelines.
Next step
Connect the read-only endpoint first and let the agent explain failures with compare_builds and the log tools. Add the retries toolset only when you want it to act. Browse the MCP server directory or read how to choose MCP servers.
Frequently asked questions
Is the Buildkite MCP server free?
Buildkite's MCP docs do not list a separate price. You need a Buildkite account, and remote OAuth usage has its own quota separate from your REST API limit.
Can I make it read-only?
Yes. Use https://mcp.buildkite.com/mcp/readonly. Buildkite notes this limits only that connection, not other tools or credentials the agent has.
Can the AI retry failed builds without full write access?
Yes. The retries toolset needs retry_builds plus read scopes, not write_builds. retry_failed_jobs retries only infrastructure failures and has a dry_run option.
Should I use the hosted or local server?
Buildkite recommends hosted OAuth for interactive AI tools. Use the local server, pinned to a version, when you need stable behavior or want to control exactly which release runs.
