The environment around your coding agent
Rules, skills, MCP servers, a code graph, and shared project knowledge. omh wires them together for each repo, so you can give Claude Code a task without assembling its environment first.
Sandboxed from your host. Commands run in a container, with host files kept out by default.
brew install mindsers/tap/omh
Open sourceMIT licensedv0.14.1 · Early release
Rules, skills, MCP servers, and hooks, in the places your agent expects.
Symbols, callers, and dependencies the agent can look up, indexed for each session.
Project decisions and findings available to the next session.
Ready for the sessions you start
Shared project knowledge. A separate Docker sandbox and code index for each session.
omh new claude
The setup you don’t have to repeat
You shouldn’t have to install graph tools, connect MCP servers, and configure project memory every time you start a repo. Run omh init. omh handles that wiring.
Code graph
Give your agents a way to find symbols, trace callers, and follow dependencies. omh configures the code graph for the project and indexes each session’s own workspace, refreshing it after every agent turn.
“Who calls this function?”
“What depends on this module?”
Shared knowledge graph
Keep decisions, conventions, and earlier findings available to the next agent. Sessions in the same repository share project knowledge, so useful discoveries don’t disappear when a session ends.
“This API still has older clients. Keep the existing response fields.”
Start the next task with the code tools and project knowledge already connected.
The rest of your working environment
Keep rules, skills, MCP servers, commands, and hooks in one place. omh puts them where the agent reads them. Export the raw config with omh eject whenever you want.
Open VS Code, Zed, Cursor, or Neovim over SSH. Inspect or change the code yourself, with the same files and dependencies your agent uses.
See your sessions, their state, and overlapping file changes in one place. Each session has its own workspace and git repository; you choose which changes to bring back.
Fewer setup decisions
omh init reads your manifests, README, and git history to configure the project. Use omh why to see why an entry was selected and how to remove it.
$ omh init … stack rust (from Cargo.toml) hooks 2 selected (6 more in your catalogue) provision rust/linker provision rust/toolchain memory 2 notes written, 0 already there …
$ omh init building omh/claude:03cc8005caa6e532 — first run only… … omh init — decided, asked nothing harness claude (found on your host) image omh/claude:03cc8005caa6e532 image omh/claude:d06de6ac8cb6e58c (this repo's toolchain) stack rust (from Cargo.toml) hooks 2 selected (6 more in your catalogue) provision rust/linker provision rust/toolchain memory 2 notes written, 0 already there graph indexing in background → omh-cache-your-project-8cd1eb33 … omh why <name> what it costs, what was considered instead, how to remove it …
Your first session
A failing test is a good place to start. Here’s the workflow, from an existing repo to a branch ready for review.
01 / Set up
In your project directory, let omh detect the stack, configure the graph tools, and build the environment. Then sign in to Claude.
omh init
omh auth claude --name personal
The first image build takes longer. Later runs use the cache.
02 / Delegate
Your configuration, code graph, and shared project knowledge are available in the session. Ask Claude to fix the test and run the suite.
omh new claude
Want to take a look yourself? Open the same environment with omh s01 attach.
03 / Review
Back in a host terminal, read the patch. When you’re happy, commit to the session branch and push it for review.
omh s01 diff -p
omh s01 commit -m "Fix parser test"
omh s01 push fix/parser-test
These are example commands. Replace s01 with your session ID from omh s. Committing runs your repo’s checks in the sandbox; your checkout changes only when you merge.
$ omh s01 diff README.md | 2 +- src/main.rs | 6 +++++- 2 files changed, 6 insertions(+), 2 deletions(-) $ omh s01 commit -m "Add a greeting" checks passed (3) committed to omh/s01 (1 commit on the branch) $ omh s s01 omh/s01 up 1 to push
A sandbox on your computer
The agent’s shell runs inside a Docker container, not directly on your host. It can install dependencies, run tests, and change files in its workspace without giving it general access to your computer.
Your home directory, personal documents, and other projects aren’t exposed by default. The session gets its project workspace and the configuration, credentials, and extra files you make available to it.
Your checkout is never mounted either. Review the agent’s changes on the host and merge only what you want.
What is shared with the sandbox?Home directory · other projects · your original checkout
Commands · dependencies · code changes · its own git repository
Before you install
omh is an early project. Here’s what works today and what to expect.
Current limitationsomh sets up the environment around Claude Code: your configuration, a code graph, shared project knowledge, and an isolated place to work. You get those tools connected for each project, plus a way to review and bring back changes.
Run omh init in each repository. omh sets up the code graph and project memory and connects them to your sessions. You can inspect its choices with omh why and change the configuration. Project knowledge is shared between sessions of the same repository.
Run omh auth claude --name personal to use Claude’s own login flow. omh stores the login locally under ~/.omh/creds/ for use by sessions. You can keep separate personal and work accounts. Account setup.
The agent’s shell runs in the container. Your home directory, other projects, and original checkout are not mounted by default. The session does receive its separate worktree, selected configuration, and account credentials. Extra untracked files such as .env.local must be explicitly listed in carry_in to be copied into it. Docker network access is unrestricted. Session details.
The agent keeps running. Use omh s01 resume to return. Stopping a session with omh s01 down keeps its worktree and branch. Review or save any work you want before removing a session. Session lifecycle.
Claude Code on Docker is the workflow verified through real work today. Adapters for other agents exist, but do not yet have the same validation. Start with Claude Code and Docker. See the current status.
omh is open source under the MIT license. You still need your own access to Claude Code; omh does not include model usage. Browse the source.
Try it on your repo
Initialize a repo. Start a session with your tools and project context in place.
You’ll need Docker running, git, a repository, and access to Claude Code. macOS and Linux · arm64 and x86_64.
omh init
omh auth claude --name personal
omh new claude
Give Claude a task when the session opens. Run omh doctor if setup needs a check.