Apron Agents — multi-agent AI coding with a human merge gate
A crew of coding agents.<br>One merge cockpit. Your call on every chunk.
Apron Agents splits a task across agents that each work in an isolated, disposable<br>sandbox git — then merges their work one reviewed chunk at a time. Your real remote is never touched.
Get started<br>View on GitHub
http://localhost:4650 — your machine
Apron Agents<br>merge cockpit<br>supervised
workspace
client.pyM
limiter.pyA
pyproject.toml
README.mdM
being changed now
A added · M modified<br>(merged so far)
dispatch
start
enter dispatches · shift+enter for a new line<br>prefer your terminal? apron task "…"
apron — 2 agents
worker-1shipped
— waiting for an issue
worker-2shipped
— waiting for an issue
awaiting review — clearance required
pattern clear — no reviews holding
merge queue
live feed
Handoff complete.<br>3 files copied into your working directory — apron stopped.<br>real remote: untouched ✓
01<br>You describe the task once.<br>↻ replay
Three things, working together
Plenty of tools run agents. Apron is about what happens between the agents and your codebase.
01<br>A disposable sandbox git
On launch, Apron creates a bare repo in a temp dir — a fully local "fake GitHub."<br>Every worker codes in its own clone and pushes branches there. One audited module is the<br>only path to git, so nothing half-baked can ever reach your real remote.
02<br>The merge cockpit
A live air-traffic-control view: watch the planner think the moment you dispatch, see<br>every file each agent is touching on a live workspace map, then review chunk-by-chunk.<br>Small, focused diffs; approve or send back with a reason; merges land one at a time with<br>tests on every candidate. Autonomous mode is the same machinery — merges on green.
03<br>Clone and run in one command
./run start sets up the environment, boots everything, and opens<br>the dashboard. When the task completes, the merged result is copied into your working<br>directory — the only bridge to reality — and the tool stops.
THE DISPOSABLE SANDBOX<br>a bare git repo in a temp dir — a fully local "fake GitHub".<br>created at launch, destroyed when apron stops. no network, no accounts.
worker-1 clone<br>an agent codes here
worker-2 clone<br>another agent, isolated
merge controller<br>merges + runs tests
push issue branches
merge approved chunks, one at a time
sandbox bare repo<br>every branch + main,<br>never on a real server
the handoff<br>merged files copied,<br>the only bridge<br>to reality
your working directory<br>the result lands here when all<br>chunks are merged. you test it.
git push<br>by you, by hand,<br>when you're happy
✕ apron never crosses this line — it cannot configure, fetch from, or push to any real remote
your real remote<br>github / gitlab — only you touch it
Bring the plan you already pay for
Workers are pluggable. Use your Claude Code subscription, your ChatGPT plan,<br>a raw API key — or try the whole flow with no account at all.
runnerpowered byneeds<br>claude-codethe claude CLI, headlessany Claude plan — whatever Claude Code already uses<br>codexthe codex CLI, headlessa ChatGPT plan or OpenAI key<br>apithe Anthropic API directlyANTHROPIC_API_KEY<br>demofake in-process agentsnothing — kick the tires account-free
Agents themselves are editable markdown files, not code —<br>Apron discovers your existing .claude/agents/ definitions read-only and hot-reloads<br>edits from its own overlay.
Get started
# clone and run — one command
$ git clone https://github.com/Ut8v/apronagents && cd apronagents
$ ./run start
# or install from PyPI
$ pip install apronagents && apron start