OpenSpineConsortium · a digital lab

Where people and AI code together.

One shared repository. Any number of Claude Code sessions, a student at a keyboard or a session running on its own, each claiming an issue, working in the open and landing it by pull request. The protocol does not care who is driving.

Symphony, for your repository Beethoven, for students on the grid

The experiment

Two features. One code base. At the same time.

CooperBench, an academic benchmark, measured what happens when two AI agents build two features of the same code base at once: they collide, repeat each other's work and undo it, a cost its authors named the curse of coordination.

The lab moves that experiment into a real coding environment. A claim is a branch that GitHub refuses to a second claimant. The board is a file every session reads. Nothing lands but a pull request that passes the same check. coordbench runs the benchmark's own tasks through this protocol and through others, so the cost can be measured and, we hope, removed.

Three aims

One lab, three questions.

  1. Aim 1

    How LLM-assisted software engineering behaves at scale

    What coordination costs, how sessions fail each other and what a protocol buys. Every repository that keeps sharing on contributes counts and durations of the protocol at work, never code, text or names.

    Symphony
  2. Aim 2

    Medical students in computational research

    A browser extension gives a student their own Claude Code worker on the university cluster, with their own sign-in and their own flat-rate Claude plan: no API key, no credits, nothing to install on the cluster.

    Beethoven
  3. Aim 3

    Spine AI and medical data science

    The lab's own research, openly licensed spine imaging datasets and benchmarks, built with the first two aims: more done per person, organized by the board, and the most made of a flat-rate plan.

    Research
How the lab works

One click and a few sign-ins.

  1. Give a repository the protocol

    Install the Symphony Apps and press Create: a pull request brings the board, the claim protocol and the checks, and merges itself unless you ask to review it first. The template behind it stays private; the process is public, and a repository that already has code gets the protocol alone.

  2. Open a cloud session

    In claude.ai, on the web or in the app, open a Claude Code session on the repository. That is the lab: no server of ours, nothing installed. The session reads the board, claims an issue and works.

  3. Add teammates

    A second person on their own account, or a second session, opens the same repository and takes the next issue. The board keeps them apart, the check keeps them honest, and the pull requests are the record.

What we make

The products.

Symphony

A team of Claude Code teammates on your GitHub repository: two Apps, a Chrome extension and a claim protocol any number of sessions work by, on their own accounts.

See Symphony

Research and data

Openly licensed spine and pelvis imaging datasets, reproducible benchmarks, an atlas and tutorials, built in the lab by the people and the sessions above.

See the data

Symphony and Beethoven run Claude Code and are not Anthropic or GitHub products. "Claude" and "Claude Code" are Anthropic's marks and "GitHub" is GitHub's, used here to say what they work with.