Will it work on my codebase?
Your language rarely decides it. Your setup does. This page shows what CodeHerder needs. It also shows what the server stores and where CodeHerder stops.
Three conditions, and you likely meet them
Your code sits in a git repository
CodeHerder drives git. Use github.com or gitlab.com. Other hosts have limits (see below).
You have a device that can already build and test it
A laptop or a small server works, as long as your normal build and test commands already run there.
Your coding agent can already work in that language
See /works-with for what each supported agent covers today.
Meet all three, and CodeHerder can drive the work. The rest of this page explains why.
Your code stays on your device
The agent reads and writes your code on a device you register. It works in its own git worktree. The CodeHerder server does not store your files or a full diff.
The server does store facts about each change. It keeps the paths of changed files and the line counts. It keeps commit subjects and SHAs. It keeps branch names and merge request links. It keeps a short clip of terminal output. It also relays live terminal output to your browser.
See /security for the full trust boundary, and /where-agents-run for how a device connects.
From your repository: a clone URL and a branch
You enter both once, the same way you'd add a repo to any tool:
ch repo create --name my-app --url git@github.com:you/my-app.git --branch mainThe --branch value is the base branch. Agents push their own task branch and open a merge request against it. They never push to the base branch. The merge stage lands the request with gh or glab. Only github.com and gitlab.com are known hosts. CodeHerder is not a git host.
See the full add-a-repo steps at /get-started.
From your device: a toolchain and a coding agent
Your toolchain, already installed
Use what you build and test with today. A device started with
--isolation isolatedruns each stage in the CodeHerder sandbox image instead. A task brief tells the agent to install a missing tool. ThechCLI updates itself.A coding agent
Use the one you would run by hand: Claude Code, Codex, or another tool from /works-with.
There are three ways to bring a device online. One starts a runner in your own AWS account with no local install. See /where-agents-run.
Your language is a tag, nothing more
CodeHerder has no list of languages. A capability is a short tag that you write.
A tag starts with a lowercase letter. After that it can use digits, hyphens and underscores. It can have one colon part, such as rails or lang:python. Tag an agent and tag a task with the same word. An agent tag only ranks the agent higher. A task tag is a hard filter: a launch config must cover it, or the task does not start. Task tags need the Starter plan or above.
See how capability matching decides placement at /how-it-works#capabilities.
Three things any stack can give a task
A command that tells the agent it passed
A test, lint or typecheck that the code stage can run and read, in any language.
A clear description of "done"
In a Story, the plan stage must fill in acceptance criteria before code starts.
A scope one person could hold in their head
One control, one endpoint, one bug. See /starter-briefs for what that looks like written out.
See /verification for how each check gates a task.
The limits, said out loud
CodeHerder is not a git host
Your code stays where it already lives. CodeHerder coordinates the work. It does not host the repository.
It needs a device you register
An agent always runs on a machine. For a trial,
ch --localruns on your laptop with no server. See /where-agents-run.Two git hosts, and token limits
Only github.com and gitlab.com are known hosts. A self-hosted GitLab is not. A private repo on another host gets no git credential. Agents use the git tokens you add with
ch device git-token create, or app tokens that the device mints. With neither, agents get no git credential.No person approves by default
No built-in stage waits for a human. To require one, turn on a required approval in branch protection on your git host. You can also add an approval gate to any stage (see /verification). See /when-agents-get-it-wrong.
No repository yet, only a prompt-to-app prototype? See /prototype-to-production for when to bring it here.
Questions people ask before connecting their first repo
Does CodeHerder support my language?
CodeHerder has no language list. The coding agent reads your code, so the real question is whether your agent works in that language. See /works-with for what each one covers.
What if my code isn't on GitHub or GitLab?
Only github.com and gitlab.com are known hosts. A self-hosted GitLab is not one. The server refuses an ssh URL to another host unless the operator allows it. A private repo there gets no git credential from the device.
Do I need to install anything special for an unusual stack?
Usually not. On a plain device the agent uses the toolchain already there. A device started with --isolation isolated runs each stage in the CodeHerder sandbox image. A workspace can also declare tools for devices to install.
Can I route a Rails task toward the agent who knows Rails?
Yes. A capability is a lowercase tag such as rails or lang:python. An agent tag ranks that agent higher. A task tag is a hard filter. If no launch config covers it, the task does not start.
Does CodeHerder read our source code to make any of this work?
The server does not store your files or a full diff. It stores change facts. These are file paths and line counts. They also include commit subjects, SHAs and branch names. It also keeps a short clip of terminal output and relays live output to your browser.
What if my repository is private, or behind a VPN?
Git traffic leaves the device over its own network. The server never proxies it. Agents do not get your SSH key or gh token. They use device git tokens that you add. An empty pool gives no git credential.
Bring your own stack.
Tell us your stack. We will help you connect your first repo and your first device.