OLIMP puts coding agents to work on your repositories from a machine you own. Add a project,
describe a task, and an agent (Claude Code, Codex or one of five others) makes the change in a git
worktree of its own; OLIMP commits it, pushes a branch and opens a pull request. Workflows chain
agents with tests, reviews and your approval, automations run them on a schedule, and all of it is
in a web panel, a REST API and an MCP server.
Features
- 🔓 Free and open source (MIT). One process on your machine or server, data in SQLite, or in
MySQL or MariaDB.
- 🔑 Agents run on your own logins: Claude Code, Codex, Copilot, Cursor, Junie, OpenCode or pi.
No API key needed.
- 🌿 Every task gets its own worktree and branch, so several agents work at once. On GitHub it
ends with a pull request, in a local repository with a commit on its branch.
- 📡 Watch a task live: the conversation, every tool call, tokens and cost. Continue it, cancel
it or run it again.
- 🔀 Workflows are graphs of agents, checks, loops, conditions, gates and git steps. Draw one in
the editor and simulate it, or let an agent plan it from a description.
- ⏰ Automations queue tasks and workflow runs on a cron schedule, laid out in a week calendar.
- 🐙 GitHub issues with a label become tasks whose pull request closes them. Issues, pull
requests, checks and merging are in the panel too.
- 🏁 Run one prompt as two or three variants, a minimal and a thorough one, compare the diffs and
keep the best.
- 🌳 An agent can split its task into subtasks, each in a worktree of its own, and merge their
work.
- 🧩 Skills from skills repositories in the skills.sh layout, your own and
the repository's.
- 🧠 An optional knowledge base per project: pages, hubs and sections that cite notes, which
agents search and extend with OLIMP's MCP tools.
- 🔌 MCP integrations per project, with their secrets stored encrypted, picked per agent step.
- 🖥️ A terminal in every task's folder and a preview of the app it serves, with a design mode
for pointing the agent at an element.
- 👥 Accounts for a team: a password and a mandatory second factor, API tokens, and each person's
own agent logins.
- 🔔 Notifications to Slack, Discord, ntfy or your own endpoint when a job finishes or waits for
you.
- 🤖 Whatever the panel does, the REST API and the MCP server do too, so your own agent can run
OLIMP.
- 🌗 Light and dark, English and Polish, and it works on a phone.
Screenshots
A task. The request, the agent's tool calls and answer, what it cost, and the branch it left.
Workflows. Make the change, run the tests, repair until they pass, commit.
Memory. A project's hubs, pages and the notes they rest on, drawn as a mountain.
Quick start
You need Node 24.2+, git and at least one agent CLI you're logged into:
Claude Code (the default),
Codex, GitHub Copilot CLI,
Cursor Agent, Junie,
OpenCode or pi. The
GitHub CLI gh is needed only for GitHub projects.
cd your-repo
npx olimp-run
This starts OLIMP at http://127.0.0.1:4850 with your repository as a project: describe a task and
send it. Add more projects from GitHub, from any local folder or as new repositories, and pick
another default agent in Settings → Agent accounts. Data stays in ~/.olimp.
Keep it running
Install it to get the olimp command. On Linux and WSL it then runs as a systemd user service that
starts with the system and comes back after a crash; the panel's version window updates it from npm:
npm install -g olimp-run
olimp service install
olimp service logs
To run the latest main instead, from a clone: pnpm install && pnpm prod
(see the guide).
Reach it from elsewhere
Without accounts OLIMP listens only on 127.0.0.1 and trusts every request that reaches it. To use
it from another machine or with a team, create the first account, put a TLS proxy (Caddy, nginx)
in front of it and set OLIMP_PUBLIC_HOST; see Accounts and tokens.
A one-command server setup is planned.
How it works
- Add a project: a GitHub repository, a local git repository or any folder.
- Describe a task in the panel, through the API or MCP, from a labelled issue or on a
schedule. Pick the agent, the model, a skill and files to attach, or run a workflow.
- The agent works in its own worktree. OLIMP runs your agent's CLI there and streams what it
does. A workflow can stop at a gate or a question until you answer.
- Review the result: the diff, the commits and the pull request. Continue the conversation,
run it again, or merge.
Agents learn a repository from its own AGENTS.md, CLAUDE.md and docs. A workflow is a graph you
draw in the panel or send as JSON, here one that repairs the change until the tests pass:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
