MCP for Code, CI, Runner Farm and QA
The CHAMPREP MCP server lets your AI assistant work with CHAMPREP Code, Runner Farm and QA on your behalf: find projects and their clone URLs, check pipelines, read job logs, look after runners and runner pools, follow QA scans and triage findings. The assistant always acts as you: your roles, your plan, your organization’s policies and (with an API key) the key’s scopes apply exactly as in the web apps.
Connect
Section titled “Connect”Connect your AI client to https://ai-api.champrep.com/mcp/v1 as described in
MCP server. Browser sign-in works for people at a
keyboard; for scripts and CI use an API key. The scopes that matter here:
| Service | Read only | Read and change |
|---|---|---|
| CHAMPREP Code | code:read_api | code:api |
| Runner Farm | runner_farm:read | runner_farm:write |
| QA | qa:read | qa:write |
Finding the right operation
Section titled “Finding the right operation”When your account can reach a large catalog (more than about 1,000
operations, which Code alone nearly is), your client sees two helper tools
instead of one tool per read: champrep_search_tools and champrep_read.
Your assistant searches with a few plain words, reads the exact input it
needs, then runs the operation. Searches such as job log, latest pipeline,
runner tags, group projects or quality gate work: every word is matched
on its own and the best matches come first. In that mode reads go through
champrep_read; with a smaller catalog every operation is listed as its own
tool. Every change always has its own named tool, so your client can ask you
before it runs.
CHAMPREP Code
Section titled “CHAMPREP Code”Projects, groups and cloning
Section titled “Projects, groups and cloning”- List the projects you can access, or every project of a group with all its subgroups. Each project comes with its HTTPS and SSH clone URLs.
- Clone a whole group. MCP cannot run
giton your machine. Ask for a clone plan instead:code_clone_group_planlists every project of the group (subgroups included) and returns ready-to-rungit clonecommands that keep the group folder layout. Folders that already exist are skipped, so you can run the list again. Run them yourself in a POSIX shell, or let a local coding agent (Claude Code, Codex, Cursor) run them in your terminal. - SSH cloning needs an SSH key on your Code profile. HTTPS cloning asks for your username and a personal access token that you create in Code: never paste a token into the chat.
Pipelines and jobs
Section titled “Pipelines and jobs”- Pipeline status per branch, the latest pipeline of a branch, the jobs of a pipeline and the test report summary.
- Job logs. Read a job’s log; for long logs the assistant reads it in pieces with a byte range, so it can jump to the end of a failed job.
- Run a new pipeline, retry or cancel a pipeline, play a manual job, retry or cancel one job.
- List and download job artifacts (small archives; large archives are downloaded from Code).
Runners
Section titled “Runners”- List your runners, a project’s runners (its own, its groups’ and the instance runners it can use) and a group’s runners, with status, online state, tags, paused state and last contact.
- See a runner’s details and the jobs it ran, and how many jobs are waiting for a runner.
- Register a new runner. The assistant opens Code’s new-runner page for the project or group. You get the runner token there: it is never shown to the assistant.
- Change a runner: add or remove tags, pause or resume it, change its description, whether it runs untagged jobs, its protection or its maximum timeout. Tags are sent as the full new list, so the assistant reads the runner first. Each change runs only after you approve it.
- Assign an existing project runner to another project you maintain.
- Delete a runner, or detach it from one project: destructive actions (see What needs your approval).
- Code decides who may change which runner: project and group runners need the Maintainer or Owner role, instance runners need a Code administrator.
- The assistant can also give you a direct link to a runner’s settings page in Code.
Merge requests and issues
Section titled “Merge requests and issues”List, read, open, update, approve and merge merge requests; comment on merge requests and issues; list, create and update issues. Approving names the target branch the assistant read, so a retargeted merge request is not approved by mistake.
Runner Farm
Section titled “Runner Farm”- Your Runner Farm accounts and workspaces, your role, and your plan’s machine classes and ceilings.
- Code host connections: list and read them; rename, pause, resume or label a connection.
- Pools: list and read them (including wait times), create a pool, rename, pause, resume or scale it (concurrency and job timeout), and restore a pool or connection from Recently deleted.
- Credits and usage: your credit balance, usage per month, class, pool or project, and the credit ledger.
- Jobs: live, held and refused jobs, and the details of one job.
- Connecting a code host and its runner token, and connector secrets, are done in the Runner Farm app: the assistant gives you a secure link and never sees the credential.
- Repositories: list connected and available repositories, activate them and sync them; CHAMPREP Code projects too.
- Scans: start a scan, follow its status, read its log (secrets are redacted) and see your scan totals.
- Findings: list findings with filters, dismiss a finding with a reason and a comment, or reopen it.
- Quality gate: the gate’s verdict for a repository or branch, with each condition and why it fails. It works for repositories from CHAMPREP Code, GitHub, GitLab and Bitbucket that are connected to QA.
- Previews: start a preview build of a branch, list previews, read their logs and build logs, extend a preview or change its visibility. Repositories that need secrets are launched from the QA app.
- Cloud IDE: start an IDE session on a repository or an empty workspace, then open it from the link to its page in the QA app; list and read your sessions. The session’s access token is never shown to the assistant.
- Starting a preview or an IDE session uses your plan’s QA allowance and runs only after you approve it.
Example prompts
Section titled “Example prompts”- “List the latest pipeline status for every project in group acme/platform.”
- “Show the log of the failed job in the last pipeline of acme/web.”
- “Retry the failed jobs of pipeline 4211 in acme/web.”
- “Which runners does acme/web use, and which ones are offline?”
- “Add the tag gpu to runner 12.” (The assistant reads the current tags and asks you to approve the new list.)
- “Pause runner 12 until tomorrow’s maintenance is done.”
- “Start a QA preview of the feature/login branch of acme/web.”
- “Open a cloud IDE session on acme/web.”
- “Register a new runner for group acme.” (You get Code’s new-runner page.)
- “Give me the clone commands for every project in group acme, including subgroups.”
- “How many jobs are waiting for a runner in acme/web right now?”
- “Scale my Runner Farm pool build-large to 20 concurrent jobs.”
- “How many Runner Farm credits did I use this month, per machine class?”
- “Start a QA scan of acme/web and tell me when it finishes.”
- “List the open high-severity QA findings in acme/web.”
- “Dismiss QA-0970 as a false positive: the key is a test fixture.”
- “Does acme/web pass its quality gate on main?”
What needs your approval
Section titled “What needs your approval”- Reads (lists, details, logs, usage) change nothing. In large catalogs
they run through
champrep_read. - Changes (run, retry or cancel pipelines and jobs, open or merge merge requests, comment, change, pause, resume or assign runners, create or scale pools, start scans, dismiss findings, start QA previews and IDE sessions) each have their own tool. Your MCP client decides how it asks you to confirm: review its approval settings before you allow changes.
- Destructive actions (delete a pipeline, delete a runner or detach it
from a project, delete a pool, stop all running Runner Farm jobs) stay hidden
until the connection allows destructive actions and the URL ends with
?destructive=true. See MCP server. - Every change is checked again by the service with your role and plan.
What is not available, and why
Section titled “What is not available, and why”| Not available through MCP | Why | Where to do it |
|---|---|---|
Running git (clone, push) on your machine | MCP tools run in CHAMPREP, not on your computer | Run the commands from code_clone_group_plan, or use a local coding agent |
| Seeing a runner token, a Runner Farm runner token or a connector secret, or changing a runner token’s expiry | Credentials must never enter a chat, prompt or log | Code’s runner pages and the Runner Farm app (the assistant gives you the link) |
| Changing a project’s CI/CD settings (for example turning instance or group runners on) | Code asks for a step-up confirmation that only you can give | The project’s CI/CD settings in Code |
| Running commands or editing files inside a cloud IDE session | Arbitrary code execution stays in the IDE you open | The QA cloud IDE |
| Creating access tokens, deploy tokens or keys, reading CI/CD variables | Credentials and secrets stay in Code | Code settings |
| Changing protections (protected branches, approval rules, members, visibility) | Code asks for a step-up confirmation | Code settings |
| Dismissing a secret finding | Confirming that a secret is revoked is your own attestation | The QA app |
| Pairing Code with Runner Farm and its connector secret | The pairing code and the secret are credentials | The Code admin area |
Which operations you see also depends on your plan, your organization’s policies, your per-service AI permissions and, for an API key, its scopes.