Home / Blog / Runners and limits
Is there an MCP server for GitHub Actions or CI so Claude can see my builds?
Yes. GitHub's MCP server has an actions toolset that is off by default. What each MCP server for GitHub Actions can read and change, and how to connect one.
Yes. GitHub’s own MCP server has an actions toolset with 4 tools that list workflow runs and jobs, read job logs and trigger a workflow, but it is not one of the 5 default toolsets, so you switch it on. Community servers cover the same ground, and ours, at https://mcp.runners.io/mcp, reads the jobs, logs and test results of a runners.io organisation with nothing to install.
- GitHub's MCP server has an actions toolset with four tools, and it is not one of the five default toolsets.
- Switch it on with --toolsets, the GITHUB_TOOLSETS variable, or the /x/actions address on GitHub's hosted server.
- A /readonly address or the --read-only flag leaves out the tools that change anything.
- runners.io's server is at https://mcp.runners.io/mcp, needs no install, and is read-only unless write access is approved and an owner switches write tools on.
- Logs are text anyone who can open a pull request can write, so give the agent read-only access.
What can GitHub’s MCP server do with Actions?
The README lists the toolsets the server ships. The default ones are context, repos, issues, pull_requests and users. Actions is a toolset of its own, described as “GitHub Actions workflows and CI/CD operations”, and it is off until you ask for it. It has four tools:
| Tool | What the README says it does |
|---|---|
| actions_list | Lists workflows, workflow runs, a run's jobs and a run's artifacts, 30 to a page by default and 100 at most. |
| actions_get | Returns details of a workflow, run or job, a run's usage, the address of a run's logs, or an artifact download. |
| actions_run_trigger | Triggers workflow actions. The README names run_workflow, which needs a ref and a workflow, and says run_id is required for every other method. |
| get_job_logs | Reads a job's log. tail_lines returns only the end, and failed_only with a run_id reads every failed job in the run. |
Each of the four lists the repo scope as its OAuth challenge scope, the scope GitHub asks a login to approve.
How do I switch the actions toolset on?
On the server you run yourself, name the toolset with the --toolsets flag or the GITHUB_TOOLSETS variable; if both are set the variable wins. On the one GitHub hosts, the address of the toolset is https://api.githubcopilot.com/mcp/x/actions, and the header X-MCP-Toolsets takes a comma-separated list, as the remote server page describes it. The local form in the README is the Docker image with the variable set. We did not run it, because it needs a token of yours:
docker run -i --rm \
-e GITHUB_PERSONAL_ACCESS_TOKEN=<your-token> \
-e GITHUB_TOOLSETS="actions" \
ghcr.io/github/github-mcp-server
Which MCP servers read GitHub Actions, and what can each one change?
Four ways to give Claude your builds, read from each project’s own page on the date this post was last checked. The last column is the one to check before you connect anything.
| Server | How you reach it | What it reads | What it can change |
|---|---|---|---|
| GitHub, hosted | https://api.githubcopilot.com/mcp/x/actions | Workflows, runs, jobs, artifacts and job logs of repositories your token or login can see. | actions_run_trigger can run a workflow and, as its run_id parameter shows, act on existing runs. Add /readonly to the address and only read tools are offered. |
| GitHub, run yourself | Docker image ghcr.io/github/github-mcp-server with GITHUB_TOOLSETS set | The same four tools. | The same tool. The --read-only flag offers read tools only. |
| A community server (ofershap/mcp-server-github-actions, MIT) | npx mcp-server-github-actions, with a token in GITHUB_TOKEN | Nine tools: list_workflows, list_runs, get_run, get_run_logs, list_artifacts and four more. | Four act on runs: rerun_workflow, rerun_failed_jobs, cancel_run and trigger_workflow. Its README asks for Actions read and write on a fine-grained token. |
| runners.io | https://mcp.runners.io/mcp, sign in when the client asks | Jobs, logs, test results, cache and spend for one organisation, as your role could see them in the dashboard. | Two tools, cancel_job and delete_cache_entry, off until an owner or admin switches them on. |
How do I add runners.io’s MCP server to Claude?
One command in Claude Code, then /mcp inside it to sign in. The server needs no install and no token: the client is sent to runners.io, you choose the organisation and press Approve. On 17 November 2026 an unauthenticated request to the address answered 401 with a header pointing at its sign-in details, which is what a client needs to start that approval.
claude mcp add --transport http runnersio https://mcp.runners.io/mcp
The same address works in Cursor, VS Code, ChatGPT and any client that speaks remote MCP over HTTP with OAuth 2.1 sign-in, and Claude’s desktop and web apps can add it from the connector directory. The page for our MCP server for GitHub Actions has the setup for each, every tool and what each role can read.
How do I stop an MCP server changing my builds?
Choose the read-only form where there is one. On GitHub’s hosted server that is the /readonly end of the address; on the one you run, the --read-only flag. A community server that has re-run and cancel tools, and whose README asks for read and write on Actions, has no documented read-only setup, so read its tool list before you connect it.
On ours, every connection is read-only unless it was approved with write access by an owner or an admin, the organisation has switched write tools on, and the agent passes confirm: true on the call. If any of those is missing the call is refused, says which, and changes nothing, and every attempt is recorded.
What should I not expect from an MCP server?
It returns what the token or the login can see, and nothing it was not given. It does not make Claude’s reading of a log right, and it does not make the log safe: a build log, a pull request title and a branch name are text that anyone who can open a pull request can write. Read whether it is safe to let an agent read CI logs before you give one write access, and how an agent reads a failed build for the routes besides a server. For the rest of what the agent can reach, the security hardening post covers the workflow side, and the runner security page covers ours.
Sources
- github/github-mcp-server (README) (github.com)
- GitHub MCP server: remote server documentation (github.com)
- ofershap/mcp-server-github-actions (README) (github.com)
- The runners.io MCP server (/docs/observability/mcp-server)
Frequently asked questions
Is there an official MCP server for GitHub Actions?
Yes. GitHub's github-mcp-server has an actions toolset with four tools that list and get workflows, runs and jobs, trigger a workflow and read job logs. It is off by default: enable it with --toolsets, GITHUB_TOOLSETS, or the /x/actions address on the hosted server.
Can an MCP server change my builds, or only read them?
Some can. GitHub's actions_run_trigger triggers workflows and acts on existing runs, and a community server can re-run and cancel runs. Use GitHub's /readonly address or --read-only flag. Ours is read-only until write access is approved and an owner or admin switches write tools on.
How do I add runners.io's MCP server to Claude Code?
Run claude mcp add --transport http runnersio https://mcp.runners.io/mcp, then /mcp inside Claude Code and sign in. There is nothing to install and no token to create; you choose the organisation and press Approve.
Which roles can read what through runners.io's MCP server?
An owner or admin can read jobs, logs, test results, cache, repositories and spend. A member can read the same except spend. The billing role can read spend and the runner sizes, but no jobs, logs or cache.