Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@BetaZero why is src/parser.js line 40 written this way?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
A developer's AI twin, small enough to commit to the repo.
BetaZero is an SFT-only twin of a single developer: a LoRA adapter over SmolLM3-3B, trained on that developer's own git history, and served to coding agents as MCP tools through the TrueForge agent harness.
No RL, no reward model, no reinforcement stage. Supervised fine-tuning only — hence BetaZero. The whole twin is one ~29 MB file in this repository, so you can clone it and run it.
Built for the Agent Harness Hackathon.
When a developer leaves, the code stays and the reasoning goes with them. Git records what changed and when; it rarely records why.
Coding agents inherit that blind spot. They read code, see something that looks wrong, and "fix" it — reverting a deliberate decision nobody wrote down.
Related MCP server: CodeGraph
BetaZero exposes a developer's recorded voice to any agent as MCP tools:
Tool | Question it answers |
| What did the developer say when this changed? |
| How would this developer describe this work? |
| Who made this call, and in which commits? |
A TrueForge agent about to refactor a file asks first, and revises its plan instead of undoing a decision it never knew about.
The twin in this repo is trained on D2, a developer with 443 commits to the public COBWEB-ca/cobweb2 repository (2008–2019). Each training row pairs a commit diff with the commit message that developer wrote.
Be clear about what that teaches. The targets are commit messages — a median of 49 characters, 85% under 80. So BetaZero learns the developer's voice and intent at commit time. It does not learn unwritten reasoning, because git history does not contain unwritten reasoning. Recovering that requires elicitation, not observation, and is out of scope here.
The corpus is derived entirely from public git history and is sanitised before
publication — see scripts/sanitize_corpus.py,
which redacts third-party contact details that appear in the upstream credits
screen and refuses to run if a redaction would alter a training label.
An adapter that ships without evidence is a claim, not a result. This repo publishes the A/B: BetaZero against the untrained base model, averaged over repeated runs and judged against the paired standard error, with the sampling noise floor stated alongside the effect.
See docs/evaluation.md — including the honest answer if
the effect turns out to be small.
Early. See the issue tracker for planned work and CONTRIBUTING.md for how changes get made.
This project is entered in the hackathon's Q Branch (code quality) track. Every substantive change reaches
mainthrough a pull request reviewed by Qodo.mainis branch-protected; there are no direct pushes.This section will link representative merged PRs, what Qodo surfaced, which findings were addressed or dismissed and why, and the follow-up review.
Apache License 2.0 — see LICENSE and NOTICE.
The cobweb2 corpus derives from a public repository that publishes no licence of its own; NOTICE records what is and is not claimed for it, and the sanitiser shows how it is prepared.
This server cannot be deployed
Persistent memory and cross-session learning for AI coding assistants (hosted remote MCP).
A shared team wiki your coding agents read and write — across every repo and every MCP client.
Codebase graphs, caller impact analysis, and recorded project context for AI coding agents.
Roadmap, tasks, releases and user feedback your coding agent reads and writes over MCP.
