Catch agent mistakes before they cost you.
A small hub on your own computer that your coding agents report to. It tells each agent which files the others hold, records which checks passed on each exact change, and keeps a history of what your agents claimed and reported.
npx agentscmStarts a hub on this computer and connects the coding agents it finds. Writes nothing to your repository.
Paste it into Claude Code, Codex or Cursor. It runs the setup, shows you each link and waits for you.
Watch the process.
Claude Code
fix/session-timeout
Tests passed · d4e5f6a
Cursor
docs/api
Working · docs/api.md
AgentSCM hub · your computer
Files held
src/session.tsClaude Codedocs/api.mdCursor
7 events kept in the history
Codex
fix/session-refresh
Handed over
Three ways it saves you work.
Agents are fast, but they repeat each other, miss broken tests and forget why a change was made. Each of those costs you tokens and time. AgentSCM deals with all three.
Spend fewer tokens.
Every repeated task is a bill you pay twice. Agents burn tokens when they redo work, edit the same file at once, or fix the same bug twice. AgentSCM tells each agent what the others are doing and what is already done, so less of your spend goes on repeats. You save the most when several agents share one repository.
How agents share workOne file, one agent. Less done twice.
Codex is editing
Workingsrc/session.tsClaude Code tried the same file
WarnedClaude Code took the next task instead
Less double work
Catch problems early.
Checks are tied to the exact change an agent made, so an old pass can’t count for a newer edit. If a test fails, you hear about it straight away, while the fix is small. Not three review rounds later, when every round costs more.
How changes are checkedCan this change go in? Not yet.
- Tests
Passed on this exact version,
Doned4e5f6a - Review
An agent reviewed
Doned4e5f6a - You
Your approval
Waiting
Change the code and the checks run again.
Keep the whole story.
Every request, change and check is saved, and old versions are kept, not written over. Weeks later you can still see who asked for what, which agent changed what, which checks ran, and who approved it. Git and GitHub still decide what merges.
How history is keptSign people out after 30 minutes. Don’t change the public API.
Sessions end after 30 minutes
MetPublic API unchanged
MetA test covers the timeout
Not yet
Don’t take our word for it. Read the source on GitHub, or run npx agentscm demo and watch three pretend agents work.
See it in two minutes.
Poster frame: the AgentSCM logo over the line “Source control for teams of coding agents.” on a dark violet background.
Up and running in four steps.
You need Node 18 or later and a Git repository. Works on Mac and Linux (Linux and Intel Macs also need Docker). Windows works too, with fewer features: you connect your agents yourself. Setup writes nothing into your repository.
Install it once.
Run this once per computer. It starts a hub on your computer and adds hooks to the coding agents it finds; some agents ask you to approve them.
npx agentscmRestart your agents.
Quit and reopen them, then open a repository and ask one:
Get started with AgentSCM.Open your hub.
See which agent is on which task, what changed and which checks passed.
npx agentscm openCome back any time.
Read a short catch-up of what happened while you were away, then hand the next task to any agent with the notes attached.
Catch-up and hand-offs
Just looking? npx agentscm demo runs three pretend agents on a throwaway hub and writes nothing to your repository.
Everything in the Community Edition.
Twelve features that follow your agents’ work from start to finish. Every one runs on your own computer.
See the work
While your agents are working.
- Who is on what
Each agent claims its task, so you can see who is doing what.
- Same-file warnings
An agent is told when another one is already editing that file.
- Catch-up
A short summary of what happened since you last looked.
- Hand-offs
Pass a task to another agent with its notes attached.
Check the work
Before a change goes in.
- Change records
Each change with its files, the chat behind it and its checks.
- Code explorer
Browse code, history and diffs beside the agent conversations.
- Pull request checklist
The steps your team asks for on every pull request, with proof.
- Goal check
An agent compares what you asked for with what was built.
Keep the history
Long after it ships.
- Written goals
Write down what you want built and the rules. Every change to them is kept as a new version.
- Change history
Requests, changes and checks saved so you can look back later.
- Approval step
Optional: a change goes in only after its checks pass and you say yes.
- Reusable workflows
Save the way your team works and use it again.
Fits what you already use.
- Your agents, connected
Claude Code, Codex, Cursor, GitHub Copilot and Gemini CLI. Set up for you on Mac and Linux.
- Keep Git and GitHub
Works beside your repository, GitHub or GitLab. Nothing to move.
- Private by default
Chats stay on your computer unless you choose to share them.
- Mac, Linux or Windows
Runs on your computer. Windows has fewer features for now.
Run it yourself today. Hosted is on the way.
Community Edition
Available now · alpha- Source code on GitHub, under FSL-1.1-ALv2
- Every feature on this page
- Runs on your computer; self-hosting on a server works but isn’t supported in the alpha yet
- Your code and chats stay with you
- Pilot it on one repository;
npx agentscm uninstallremoves it
Hosted
PlannedWe run the hub for you, so there is nothing to install on a server. It is being built now at agentscm.ai.
Visit agentscm.aiQuestions people ask.
What does alpha mean?
Good for a pilot on one repository, not yet for work you depend on. Commands and storage can change between releases, so read the release notes. Edit warnings are a nudge, not a lock. An agent’s “finished” and an agent review are reports, not proof. The approval step and written goals are experimental.
Where does it run?
Apple Silicon Macs run the hub natively. Linux and Intel Macs need Docker or an existing hub. Windows has its own build: run npx agentscm in PowerShell. It runs the hub and the edit warnings, but doesn’t set up your agents for you yet, and some features are Mac and Linux only; see the Windows notes. Teammates can join a hub that is already running with npx agentscm --hub-url. Running the hub on a shared server works but isn’t supported in the alpha; see the server guide.
What does setup change on my computer?
It installs the agentscm program, starts a hub on your computer and adds hooks to the coding agents it finds. Some agents ask you to approve those hooks. It writes nothing into your repository and never blocks a Git push. npx agentscm uninstall removes what setup added.
Do I have to stop using Git or GitHub?
No. AgentSCM works beside your Git repository and GitHub or GitLab. They still decide what gets merged; AgentSCM never merges for you.
Which coding agents does it work with?
Setup connects Claude Code, Codex, Cursor, GitHub Copilot and Gemini CLI. They don’t all support the same things yet; the agent guide shows what each one does.
Does my code or chat leave my computer?
The hub runs on your computer and keeps task notes and who is editing what. It doesn’t keep your chats unless you turn that on for a repository. If you share chats with your team, secrets are removed on a best-effort basis and shared copies are kept for 30 days by default. This website uses analytics; the privacy notice explains what it collects.
What is agentscm.ai?
The hosted version we are building, so teams don’t have to run a hub themselves. It isn’t available yet and has no date. To run AgentSCM today, use the Community Edition.
What license is it under?
The Community Edition is published under the FSL-1.1-ALv2 license: you can use it, read it and change it, but not sell it as a competing service. Each release becomes Apache 2.0 two years after it ships.
Try it in one command.
On your own computer, in one line. Remove it any time with npx agentscm uninstall.
npx agentscm