The AI team workspace. Build and manage a team of AI agents in files you own. Free and source-available. Works with the Claude or ChatGPT subscription you may already have. Built by someone running their own business, for people running their own.
Download for macOS or Windows · guided setup; after it, your day happens in Rundock.
You run content, ops, sales, admin, and research. When you started, there was nobody else, so the work fell to you. A single chatbot is a single assistant. An agent platform built for developers assumes you can write code. Rundock gives you a team you can actually manage: an org chart of named specialists, parallel conversations you can watch side by side, and delegation that happens in front of you. One beta user described it as having a virtual team of highly paid experts, running in parallel. That is the experience.
Everything the team learns lives in the workspace as files you own. Skills and knowledge carry forward instead of resetting with every session.
Built by Liam Darmody. Learn more at rundock.ai.
Star this repo if Rundock is useful to you. It is the simplest way to support the project and helps other people running their own businesses find it. Add a star ->
The org chart is the product. Show someone yours.
Review your agent's work where it lives. Select any passage in a file an agent produced and leave a comment. It proposes an inline edit you accept or reject, stored in the file as plain text you can read in any editor.
Find anything with one shortcut. Press Cmd+K to search file contents, conversation messages, and agent and skill names at once, then open the result at the right place. Lists group your conversations by project or area of focus.
Your boards open as real columns. An Obsidian Kanban board opens as a column board you drag cards across, and writes back the same markdown so it stays interchangeable with Obsidian.
Pin the files you live in. Pin a file from its view or the sidebar and it sits in your Pins list, one click away, with its folder underneath. Pins are kept with the workspace, so they follow it to another machine when the folder syncs; they stay out of git, so a shared repository keeps each person's own list.
See how your workspace connects. Map draws every file as a node and every link between files as a line. Notes that link to each other cluster together, recently changed files glow warmer, and clicking a node opens the file.
Schedule a skill to a cadence, through a form, and it runs without you watching. Pick an agent, a skill, and how often, and Rundock runs it on its own server: a morning briefing, an end-of-day sync, a weekly digest. See what a run actually did from its own screen, run one now to try it, or stop one that's misbehaving.
Add a team or a new view from one link. Paste a GitHub link into Settings, Packages and Rundock shows what the package holds before anything is written: its agents, skills, routines and any extension. An extension opens a kind of file in its own view, built from Rundock's own components and run in a sandbox. Browse what is available in the Directory.
Rundock is two things at once. For you, it is the visual workspace: an org chart, parallel conversations, delegation you can watch happen. Underneath, it is a runtime-agnostic orchestration layer. It catches requests, routes them to the right agent, hands context between them, and keeps the team coordinated. The execution engine is pluggable: Claude Code and Codex today, more later. Your agents are plain markdown you own, so the same team runs on whichever engine you choose, switchable without a rebuild. Rundock owns the coordination, the interface, and the portability. The engine underneath can change. The layer you built stays yours.
Five ideas shape every decision in Rundock: local-first, markdown all the way down, the human leads, the team is the unit of value, and built from real use.
Read the full version at docs.rundock.ai/principles.
You do not need to write code. You need a subscription that can run agents (Claude Pro or Max; a ChatGPT plan works too, see Runtimes) and a folder to call your workspace. Rundock's first-run wizard walks you through installing Claude Code (one command to paste) and connecting your Claude account.
-
Get a Claude Pro or Max subscription. Claude Code runs your team by default; Rundock's wizard installs and signs you in.
-
Download Rundock for your platform, from rundock.ai/download or the releases page:
Platform Download macOS (Apple Silicon or Intel) Universal .dmgWindows 10 / 11 (64-bit) .exeinstallerLinux Build from source On Windows the installer is unsigned for now, so SmartScreen shows a one-time prompt: click More info, then Run anyway.
-
Open Rundock. The first-run wizard installs and signs you into Claude Code, then Doc, the built-in guide, walks you through choosing a workspace and creating your first agents.
Agents run on Claude Code by default. Any specialist can instead run on Codex, the official OpenAI CLI, using a ChatGPT plan: add runtime: codex to that agent's file (or ask Doc to do it) and leave the model field out so your account's default model applies. Codex agents converse, resume threads, take delegated work from your orchestrator, and review files like any other teammate; they run inside Codex's built-in sandbox, and in Code mode Rundock judges their commands by the same rules as Claude Code's. The workspace orchestrator always runs on Claude Code.
Codex setup is currently manual (two commands in a terminal): npm install -g @openai/codex, then codex login. Rundock's Settings show each runtime's status. On Windows, one config line gives Codex agents direct sandboxed file writes; without it, their writes arrive as approval cards. Details at docs.rundock.ai/concepts/runtimes.
For Linux, or anyone who wants to run Rundock from source. This is also the contributor path.
Requirements: Node.js 22+ and Claude Code authenticated (claude --version should work in your terminal).
git clone https://github.com/liamdarmody/rundock.git
cd rundock
npm install
npm startOpen Rundock from the link it prints in the terminal (Rundock is running: http://localhost:3000/#c=…). Each browser needs the link once: after that, new tabs and restarts open Rundock as before. To open a specific workspace directly:
WORKSPACE=/path/to/your/folder npm startTo pull updates later, run npm run update in the install directory.
- ARCHITECTURE.md: the process model, the desktop app and its advanced settings, the workspace directory, and codebase structure.
- CLIENT-ARCHITECTURE.md: how
public/works. No build step, classic scripts, why modules publish onto the global object, and which settings are load-bearing. - AGENTS.md: the agent file format reference. Frontmatter fields, the markdown body, workspace modes, and a complete example.
- SKILLS.md: the skill file format, discovery, and the assignment model.
- ROUTINES.md: the schedule format, scheduler behaviour, and where output goes.
- PACKAGES.md: what a package carries, what the install offer says about it, and the receipt an install leaves.
- EXTENSION-HOST.md: the extension host contract. What a mounted extension can reach, what it cannot, and how each line is enforced.
- RUNDOCK-UI.md: the component library Rundock puts into every extension view.
- DESIGN.md: the design standard. Every token with its purpose and usage rule, the two rules that decide a radius, and what is deliberately left untokenised.
- CONTRIBUTING.md: dev setup, code structure, conventions, changelog standards.
- CHANGELOG.md: release history.
- LICENSE: PolyForm Perimeter 1.0.0.
The entire stack runs on your machine. Rundock never sends your files, your agents, or your conversations anywhere. Rundock itself contacts only GitHub: the desktop app checks there for its own updates, and Rundock fetches a package from there when you add it, check it for updates or update it. None of your files, agents or conversations go with those requests. Everything else is your runtime and the connectors you choose: Claude Code talks to Anthropic's API, Codex (if you use it) talks to OpenAI's, and any connector you add reaches its own service. The runtime sends only the active conversation, processed under your own subscription. There is no cloud service, no account to create, no server-side database, no telemetry. Sign-in and keys are managed by each runtime's CLI, and Rundock never reads their contents: it learns that Claude Code is signed in from how its turns go, and that Codex is signed in from whether its credentials file exists.
Inside the app, every action that matters asks first. An agent that wants to run a command or change a file outside its permissions raises a card you allow once, always allow, or deny. In Notes mode, a connector action whose name says it deletes, sends, replies or forwards asks every time, with no always allow, because it can't be called back, while reading through a connector, such as suggesting a meeting time, doesn't ask. An email or calendar card shows who it goes to and what it says: recipients, subject and message, or an event's title, times and attendees. On a Mac, a command that reads or changes a password or key in your Keychain asks every time in both modes, with no always allow. Settings, Permissions lists everything you have always allowed, and takes any of it back. Notes mode suits a workspace of documents; Code mode lets agents run everyday development commands without asking, while commands that can't be undone still ask every time. Details at docs.rundock.ai/concepts/permissions.
PolyForm Perimeter 1.0.0. Fork it, audit it, learn from it. The one thing you cannot do is use the source to build a product that competes with Rundock. See LICENSE for the full terms.
Early access. Bugs and ideas welcome at github.com/liamdarmody/rundock/issues.