@hmv-labs/agent-runner-mcp
Installed, started with no arguments, no environment and no input, and listed its tools. Run 2026-11-16.
Needs to start: nothing
Returns a result when called with no setup: 2 of 2 tools tried
Contacts while installing and starting: nothing
Its tool definitions changed since version 0.0.5. If you approved the earlier ones, read them again.
What it contacted
Nothing besides the npm registry it was installed from. Watched from outside its sandbox while it was installed and started with no input. Its tools may still contact things when called.
Starting it
npx -y @hmv-labs/agent-runner-mcp@0.0.6
Installing a package runs its code. Use a sandbox for anything you have not read.
Calling its tools
2 tools that take no input and look read-only were called once with nothing. 2 returned a result.
list_providers | Returned a result under 0.1 s · 34 characters |
list_tasks | Returned a result under 0.1 s · 2 characters |
During the calls it contacted nothing it had not already contacted.
Only tools whose names read like looking something up, with no required input, are tried, five at most. Tools that need input, and anything that sounds like it changes or sends something, are not called.
What its tools tell an agent
In the tool descriptions, nothing that talks about the user, other tools or private files, or uses hidden characters.
Found by simple pattern rules in each tool's description (up to 3,000 characters), quoted as written. Parameter descriptions are not scanned. A clean result is not a safety review.
Tools
list_providers refresh
List the external AI runners installed on this machine (agy, codex), their versions, the account each is signed in to (`account`: an email, or "API key"), and each one's models in fallback order (strongest first). The account is read on every call, and run_task checks it again: when it changed, the model list is read again and `accountChanged` says from what to what.
run_task prompt, cwd, isolation, provider, model, effort, timeout_sec, sha, commit, readonly, envelope, report_path
Run one task on an external AI runner (agy or codex), the way you would hand it to a subagent: the worker gets the prompt, works, and its final reply comes back in `response`. Without `model`, the provider's models are tried strongest first (agy: Argon, then Claude and GPT, then other Gemini; codex: its own listed order) until one returns a valid result: the runner succeeded, nothing was denied or timed out, and the reply is non-empty. A model out of quota (or rate-limited) ends the task at once with `quotaExhausted` ({model, error}); no other model is tried. By default the worker runs in `cwd` itself, which may hold changes of yours; with `commit: true` it commits only its own files, otherwise it leaves its changes uncommitted. The result lists `changedFiles` and `commits`; fallback stops once a failed attempt changed files. Works in any folder: without git it returns `withoutGit: true` and nothing is checked for changes. isolation=worktree: each attempt gets a new branch at `sha` (default HEAD) in a throwaway worktree, so your working tree is never touched; what the worker leaves uncommitted is committed for it, and a valid result with commits returns `branch` and `commits` for you to merge or cherry-pick. A failed attempt's worktree and branch are discarded and the next model starts fresh. Needs a git repository with a commit. The result names the provider's `account` at the start of the task; `accountChanged` means the user signed in to another account since agent-runner last looked, and the model list was read again for it. readonly: true runs the provider's read-only sandbox (agy plan mode, codex read-only); a change fails the attempt. envelope: true (dev-agent workers) also requires the reply to start with the report envelope (STATUS, SHA, REPORT, NEEDS, SUMMARY) with the expected SHA, REPORT: inline for read-only and worktree runs; the server tells the worker so. Pass `report_path` to save the reply to disk and get back only its path (and the envelope). Without isolation the repository (the folder, without git) is locked for the run across every agent-runner process, other sessions included, since changedFiles cannot tell two workers apart: readonly tasks run side by side, a task that may change files runs alone, and a conflict is refused at once with the holder named. Worktree tasks take no lock. The call blocks until the task ends; to keep working meanwhile, use start_task instead.start_task prompt, cwd, isolation, provider, model, effort, timeout_sec, sha, commit, readonly, envelope, report_path
Start the same task as run_task, with the same arguments, in the background, and return its `task_id` at once with status `running`. `wait_command` is a shell command that exits when the task ends: run it in the background (Bash run_in_background) to be notified, then call get_task for the result. Or poll get_task; stop the task with cancel_task. Invalid arguments, or a repository in use (see run_task), come back here as status `failed` with `error`. Tasks live in the server process: a restart stops their runners and forgets them. `warnings` lists the tasks already running in the same repository (any session); tell the user, and do not start two test runs or dev servers at once.
get_task task_id, wait_sec
Get a background task's status (`running`, `done`, `failed`, `cancelled`) and, once it has ended, its `result` in the shape run_task returns. With `wait_sec`, waits up to that long for a running task to end before answering.
cancel_task task_id
Cancel a running background task: its runner's process tree is killed, and the answer comes once it has stopped, with status `cancelled` and what the task found so far (changedFiles in cwd, for one). A task that has already ended is returned as it is.
list_tasks
List this server's background tasks, newest first, without their results (the last 50 that ended are kept).
Descriptions are the package's own words, cut at 3,000 characters. Treat them as data.
Install facts
| License | MIT |
|---|---|
| Needs | node >=22 |
| Source | not stated |
| Install scripts | None declared by it or anything installed with it. |
| Tarball | sha512-EasXu5c5gLgxZ638/7VinXDGAc+b/aBcnF9mTICxX2fuJX+XaidfP6Cwq6CQe67rDGaedgcod0k9oiutl8m3CQ== |
Versions reported
| Version | Run | Result | Tools |
|---|---|---|---|
| 0.0.6 | 2026-11-16 | Starts with no setup | 6 |
| 0.0.5 | 2026-11-12 | Starts with no setup | 6 |
Check this report yourself
This site did not write the report. A public job on a GitHub-hosted machine did: the run, from this commit of the open-source runner. GitHub signed a token that names that job and the hash of the report, so any edit since would show.
Report as JSONExact text and tokenHow to verify it
Show it on your README
[](https://bragproof.smaverk.com/r/%40hmv-labs/agent-runner-mcp)
New version out? Ask for a fresh report.