Agetor logo

The Harness
Orchestrator

Run multiple coding agents. On every repo. At once.

A harness is the runtime that drives a coding agent — the loop, tool dispatch, and context around the model. Claude Code is a harness; Codex and more are coming soon. Agetor is the kanban control plane that runs them in parallel — including multiple aliases of the same harness, each with its own account or binary — isolates every task in its own git worktree, and surfaces approvals and clarifying questions as structured cards.

macOS
Local-first
Multi-harness
Open source
MIT licensed
Agetor app preview
Features

Built for running harnesses at scale

Everything you need to coordinate multiple coding agents at once — without juggling shells, branches, and accounts by hand.

Harnesses, not just agents
A harness is the runtime that drives a coding agent — the loop, tool dispatch, and context around the model. Claude Code is a harness; Codex and more are coming soon. Run multiple harnesses, or multiple aliases of the same one (second account, custom binary, different env), side-by-side with dedicated $HOME so logins, history, and config never collide.
Per-task git worktrees
Every task runs on its own branch in a dedicated worktree off a pinned base ref. Two harnesses can work the same repo without stepping on each other — re-runs always start from the same commit.
Persistent harness sessions
Each harness runs in the execution model it was built for — long-lived sessions for interactive agents, one-shot spawns for batch ones. Follow up on a task without losing context. Assistant text, thinking, tool calls, and results render with their own UI when the harness emits them.
Approvals as structured UI
Permission prompts and clarifying questions surface as cards in the run panel — radios, checkboxes, free-text — instead of getting buried in the harness's terminal. Approve once, deny, or remember the decision for the task.
Real-time updates
Live output streams into the run panel as the harness produces it. Native notifications and toasts fire the moment a run finishes, fails, or blocks waiting on input — so you can keep multiple harnesses busy without babysitting any one of them.
Reproducible re-runs
Run history, exit codes, durations, and full transcripts persist per task. Cancel keeps the session alive for iteration; delete tears down the worktree, branch, and session cleanly.
How we run Claude

Interactive mode only. No claude -p, no Agent SDK.

Agetor drives Claude through its interactive REPL — the same path you take when running the CLI by hand. We don't call claude -p (print mode) and we don't use the Claude Agent SDK. How that gets billed depends on Anthropic's current terms; we just stay on the path you'd take yourself.

Interactive Claude REPL
Same execution path as the CLI
claude -p / print mode
Claude Agent SDK
How it works

From prompt to merged branch

Three steps. No setup wizard, no cloud relay.

01

Pick a repo, pick a harness

Choose any local folder and any harness — built-in or one you defined. Write a prompt. If the folder is a git repo, you get worktree isolation for free.

02

Agetor spawns the harness

A dedicated branch and worktree get created off your base. The harness runs there in its own environment, with the right execution model for the agent it wraps. Output streams back live as it's produced.

03

Review, approve, follow up

Tool approvals and clarifying questions surface as structured cards in the run panel. Send follow-up turns without losing the conversation. Re-run any task to a clean base state.

Why Agetor

The harness orchestrator, not another agent

Agetor doesn't replace your agents — it runs them. No proprietary model, no SDK lock-in, no guard-rails on top of the ones the agents already have.

Local by default

Everything — tasks, runs, transcripts, approvals — lives on your machine under ~/.agetor. No relay, no telemetry, no account.

Isolation where it matters

Per-task git worktrees and per-harness $HOME keep parallel runs from interfering with each other. Agents act with your shell privileges — the same as running the CLI yourself — so what you trust stays consistent end to end.

Composable, not a platform

Built on the tools you already use — git, your shell, your editor. Add a new harness kind by extending two files; no plugins, no SDK lock-in.

Designed for parallel harnesses

Run as many tasks as you want in parallel — across multiple repos and multiple harnesses. Each one is independently cancellable, resumable, and inspectable.

Install

Up and running in under a minute

Download the app for macOS, or build it from source yourself.

Recommended

Download for macOS

One-click install. The signed .dmg has everything bundled — no build tools required.

Download for macOS

For Apple Silicon

Build from source

For Linux, custom forks, or if you just prefer it. Requires Bun, git, tmux, and at least one harness CLI on your PATH.

~/projectsbash
$git clone https://github.com/alamops/agetor.git
$cd agetor && bun install
$bun run build && bun run dev

Don't have a harness yet? npm i -g @anthropic-ai/claude-code. Codex and more harnesses are coming soon.

Stop tab-switching between harnesses.

Clone the orchestrator, point it at a folder, and ship something useful before your coffee's cold.