corkboard: every agent you're running, on one board

corkboard is a workspace for running many Claude Code agents at once. Each session is a card on the project's board. Agents claim cards, work in git worktrees of their own and hand their work back for review; one look at the board says which are working, which are waiting on you and which are ready to merge. It is our own product, and it is open source.

17 daysfrom the first commit to a signed release for every desktop.

243commits, about 70,000 lines of TypeScript, Lua and Nix.

4places to use it: Neovim, the desktop app, the browser and the phone.

The corkboard board for a project called trailhead: a "your turn" row with two cards Claude has finished, each marked low risk with a summary, above backlog, ready, doing, review and done columns. — Plate 1. The board. Two agents have finished and are waiting on a review, so their cards sit in the “your turn” row at the top. A demo project; the agents are real Claude Code sessions.

A front end built around the review

Running one agent is easy; running six means losing track of which one needs you. The board answers that first. A card shows who is working on it, whether it is running, blocked on a question or done, what it has cost and a one-line summary of what changed. Everything is a keystroke away: a new card, a new session, a review.

Sessions beside their card: Opening a card opens its agent: the live Claude Code session in the middle, the card and the rest of the queue beside it. Terminals, diffs and transcripts are ordinary Neovim windows, so they split, search and yank like any other buffer.
Review in one screen: When an agent submits, a second agent reads the diff and grades the risk. The review screen puts that grade, the agent's own summary, the branch and the size of the change together, with approve-and-merge and send-back on single keys.
The same board everywhere: The interface runs in Neovim, in an Electron desktop app, in the browser and in a native Android app, all reading the same synced board. A session started on the laptop can be answered from the phone.

A live Claude Code session inside corkboard: a test diff, a bun test and curl run, a git commit and a submit_for_review call, with the card and the ready queue in a panel on the right. — Plate 2. An agent at work on its card: the tests it changed, the checks it ran against the live server, its commit, and the hand-off for review.

The review screen for a card: branch card/task-v919 into main, one commit, three files, +75 and −15 lines, a reviewer verdict of low risk and the agent’s summary of the change. — Plate 3. Reviewing the same card. The reviewer agent's grade sits above the working agent's summary, which flags the one decision it made on its own.

Installable on day one

A tool like this is only useful if it installs in a minute. We wrote the landing page at corkboard.fyi and the release pipeline behind it: a signed and notarised DMG for Apple silicon and Intel with a Homebrew cask, a Windows installer, AppImage, deb and rpm packages for x64 and arm64 with our own apt and dnf repositories, an AUR package and a Nix flake. The desktop app updates itself. Five release candidates went out between 3 and 9 November.

The corkboard.fyi landing page: the headline “Every agent you’re running, pinned to one board.”, a short description, a download button for Linux and a table of downloads for macOS, Windows and Linux. — Plate 4. corkboard.fyi. The page picks the right download for the visitor's machine and lists the rest beneath it.

Sync, and a fleet of machines for the agents

Boards sync between devices end to end encrypted. The relay is a Cloudflare Worker that only ever stores ciphertext, and anyone can host their own; devices pair with each other and a recovery phrase brings an account back. Agents can run on the laptop or away from it: each gets a NixOS virtual machine, on the laptop or on a cloud server, and the fleet view shows every machine, its memory and the sessions running on it.

The Fleet view: a cloud server and a laptop, each with a memory bar and the virtual machines running on it, and the Claude Code sessions in each machine with their status. — Plate 5. The fleet: a cloud server and a laptop, the agent machines on each and the sessions running in them. These are our own machines.

Where the numbers come from

Days, commits and lines are counted from the public repository: the first commit was on 23 October 2026 and release candidate 9 shipped on 9 November. The board, session and review screenshots show a small demo project, worked on by real Claude Code sessions; the fleet screenshot is of our own machines.

Building something like this?

We can design the interface, write the app on every platform it needs, and build the infrastructure it runs on.

Start a conversation or book 30 minutes