Scholion

σχόλιον (scholion): a note written in the margin of a manuscript. Scholion keeps notes in the margins of your code, so your AI agents stop forgetting.

A short story

Monday. Claude Code spends an hour on a flaky network call. First it tries the obvious fix, a retry library, and finds out the library isn't installed in production. Then it finds the real fix, gets halfway, and runs out of context.

Tuesday. Codex picks the job up. Git shows it a diff and nothing else. So it reads the whole repo again, reaches for the same retry library, hits the same wall, and burns the same hour.

Tuesday, with Scholion. Codex opens the task and gets a one-page brief: the goal, what's done, "tried the retry library: not installed in prod, don't", and the three files that matter. It starts where Claude stopped. Gemini reviews the work, and the reviewer merges it.

Git remembers what changed. Scholion also remembers why, and what didn't work, and hands it to whichever agent comes next.

Watch the recorded run

What you get

• Notes in the margin. Agents leave one-line notes (decisions, dead ends) pinned to the code they touched. When any agent opens that file later, the note comes with it.
• A brief instead of a re-read. Every agent starts from a small brief, sized to a token budget: the goal, the history that matters, and the code to look at.
• One backlog for all your agents. Claude Code, Codex, Gemini, Cursor and Grok share it. Each takes the work it's best at, another agent reviews it, and the reviewer merges.
• A dashboard at http://localhost:7420/ui to watch them work.

It runs on your machine. No Docker and no account, and it never sees your model keys.

Does it help?

Measured so far, including where it didn't help:

• On a small Python repo, Claude Code needed 12.9 % fewer turns per task at the same cost (benchmark).
• When one agent's lessons were passed on to the next, cost per solved task dropped by about 42 %, and 30 of 30 tasks were solved instead of 18. Scholion did as well as a hand-written NOTES.md, without anyone having to write it.
• Not measured yet: large repositories.

Try it

The easy way: ask Claude. Add the Scholion plugin to Claude Code once:

claude plugin marketplace add vinhqbui/scholion
claude plugin install scholion@scholion

Then open Claude Code in your repository and say "set up Scholion for this repo" (or run /scholion:setup). Claude installs it, connects your agents and checks that everything works. /scholion:team sets up several agents to share one backlog, and /scholion:doctor fixes a setup that isn't working.

By hand (Linux, macOS on Apple silicon, or Windows through WSL):

curl -fsSL https://raw.githubusercontent.com/vinhqbui/scholion/main/scripts/install.sh | sh
cd your-repo && av init                       # starts Scholion, adds this repo, connects Claude Code
av init --for codex --for gemini --for cursor # the same for other agents

Restart your agent, then ask it to "start an av task to …". Open http://localhost:7420/ui to watch.

⭐ If it helps, star it

Scholion is a small open-source project. If it saved your agents a re-read, a star is how other people find it. Bugs and ideas are welcome in issues.

Learn more

• Reference: every command, configuration, the Docker stack, design, architecture and the full measurements.
• Changelog · Contributing · Benchmarks
• The CLI is scholion, with the short alias av.

How it was built. Scholion is vibe-coded: AI coding agents (mostly Claude Code) wrote almost all of the code, tests and docs, with one human steering and reviewing. Treat it as an experiment, not as audited infrastructure. Some of the design borrows ideas from Anthropic's public article Redesigning source control. Scholion is an independent project, not affiliated with or endorsed by Anthropic.

Editions. This repository is the free Community edition (MIT). It works with any number of agents, with no seat limit. A hosted Commercial edition is planned but not available.