Agentarivm
Coding agents on your machine.
Your phone in control.
Agentarivm runs coding-agent sessions — Claude Code, Codex, and whatever comes next — on a dedicated Mac or Linux box you own. Each session gets its own git branch and worktree, works toward a pull request, and reports progress to a mobile app — so you can start work, watch it, and merge it from wherever you are.
$ curl -fsSL https://agentarivm.com/install.sh | shmacOS (Apple Silicon & Intel) · Linux (x64 & arm64) · Windows via WSL2
The loop
One manager process on your machine owns everything privileged. The phone only ever talks to the manager.
Phone asks
Pick a registered repository in the app, describe the task, choose model and effort. Or label a GitHub issue and let it start on its own.
Manager prepares
The manager cuts a fresh branch from the default branch and creates a dedicated git worktree — so parallel sessions never step on each other.
The agent works
A real agent session — Claude Code or Codex, your choice per task — runs in the worktree, commits as it goes, and opens a draft pull request when it is done.
You review & merge
Watch the progress timeline, changed files and CI checks in the app; mark the PR ready and merge it — after a fresh server-side validation.
Built like a control plane, because it is one
A system that starts coding agents on your machine is remote code execution by design. Agentarivm treats it that way.
Self-hosted, no cloud
Everything — the manager, the database, the worktrees, the agents — lives on a machine you own. There is no hosted service and no usage telemetry — the manager sends us nothing but its daily license check.
Credentials never leave the machine
GitHub and agent-engine credentials stay on the manager. The phone is an unprivileged client: it never holds a GitHub token, an engine token, or shell access.
Mutual TLS pairing
Pairing mints a one-time code on the machine itself. Each device gets its own client certificate from the manager's private CA, and its token is bound to that certificate — copied to another phone, it is useless.
A branch and worktree per session
Every session gets its own branch cut from the default branch and its own worktree. Run several sessions against one repository without collisions.
GitHub issue automation
Label an issue agent-task + agent:ready and the
manager starts a session for it — but only when the label was applied by
someone on your allowlist. Off by default.
Review before anything lands
Agents open draft pull requests and are forbidden from merging. Merging happens from the app, after the manager re-reads mergeability and checks from GitHub.
The app, on every screen
The manager runs headless on your machine; the app is how you drive it — create tasks, watch the agents work, review and merge. Pair as many devices as you like; each gets its own client certificate.
Runs where you run
One dedicated, always-on machine is all Agentarivm needs.