COMU — AI software engineering in VS Code

Version 0.3.0 · Preview. Model-agnostic, bring your own key, MIT licensed.

COMU is an AI coding agent that runs inside VS Code. It reads a codebase, plans a change, edits files, runs the test suite, and asks before doing anything it cannot undo. You supply the model and the API key; COMU never proxies or resells inference.

Read this before installing. 0.3.0 is a preview. It has completed real tasks against a real model, and it has known defects that will affect you. They are listed under Known limitations, not buried. If you are looking for something dependable for daily work, this is not that yet.

Security notice for anyone running 0.2.4 or earlier

Uninstall it. Versions up to and including 0.2.4 start a local HTTP server that binds every network interface, allows every web origin, and requires no authentication. That server can create tasks which read and write files and run shell commands. Any web page you visited while VS Code was open could drive it, and on a shared or public network so could another machine.

0.3.0 fixes this with four independent controls, each verified against the packaged build:

Upgrading does not disinfect anything that already happened. If you had 0.2.4 installed on an untrusted network, treat the workspaces you opened as having been reachable.

What works

These are exercised by the automated benchmark against NVIDIA Nemotron, graded by running the code rather than by asking the agent whether it succeeded:

• Single-file bug fixes. Read the failing test, find the defect, fix it, confirm the suite is green. This is the most reliable thing COMU does.
• Small feature additions and multi-file renames, with mixed results — see the limitations.
• Approval before irreversible actions. File writes and commands stop for a human at the default ask autonomy. git push always asks, at every autonomy level, and cannot be pre-authorised.
• Cancellation. Stop aborts the run; every tool takes a required AbortSignal and a conformance suite asserts that each one refuses an aborted signal and writes nothing.
• Python and TypeScript, including repositories containing both.

Known limitations

Honest list. Each is reproducible and each has a benchmark fixture.

• A correct result can be reported as a failure in a polyglot repository. COMU detects the project type from package.json before pyproject.toml. A Python project carrying a package.json for unrelated front-end tooling is treated as a Node project, the wrong toolchain runs, and verification fails even though the change is correct and the tests pass. Measured at 5 out of 5 attempts on the fixture built for it. If you work in a mixed repository, expect COMU to tell you it failed when it did not.
• Long read-only questions are cut short. The repair budget is measured from the start of the task rather than from the start of repair, so a task older than three minutes can be terminated with REPAIR_TIMEOUT — including "explain this codebase" questions, where there is nothing to repair. Large repositories are the most affected.
• Multi-file refactors are not dependable. One observed run renamed a definition, correctly identified the five call sites, and then stopped without updating them, leaving the package unimportable. Review the diff before accepting.
• The agent sometimes writes scratch files into your workspace, such as an ad-hoc test script next to your source. Check the Changes tab.
• Slow models can exhaust the default request timeout. The default is two minutes per model request, which a large model on a large prompt can exceed.
• pnpm test does not pass at the repository root. Individual suites pass; the aggregate task does not. This is a known debt, recorded rather than hidden.
• No published benchmark numbers yet. A full baseline is being measured now. Nothing in this README quotes a success rate, because an honest one does not exist yet.

Not verified

Stated plainly so you know what has and has not been checked:

• The packaging, the loopback binding, the token enforcement and the CORS policy in the table above were verified against this exact packaged build.
• An end-to-end task driven from the VS Code interface in this build has not been re-verified since packaging, because the benchmark was holding the model provider. The runtime, the approval gate and the tool layer are covered by automated tests; the interface path is not covered by this statement.

Getting started

1. Install

code --install-extension comu-ai-0.3.0.vsix

2. Open it. Ctrl+Shift+P → COMU: Open Chat. The sidebar appears and the agent runtime starts in the background on 127.0.0.1:3456, authenticated with a token generated for that session.

3. Add a key. Click ⚙ Providers and paste a key for one of:

Keys are stored in VS Code SecretStorage (Windows DPAPI, macOS Keychain, Linux Secret Service). They are never written to the webview, the DOM, logs, or disk in plain text.

Modes and autonomy

Modes — Auto picks one for you; Agent edits and runs commands; Plan designs without touching files; Ask investigates read-only; Chat just talks.

Autonomy — ask (default) stops for approval before each write or command; auto runs without stopping, except git push, which always asks; readonly never writes or executes, whatever the mode says.

An approval shows the actual diff or the exact command. An approval that expires counts as a denial, never as consent, and with nobody watching the panel the answer is no.

Commands

Settings

Telemetry

None. COMU sends nothing anywhere except your prompts and code to the model provider you configure, over TLS, directly from your machine. There is no analytics service, no crash reporter and no usage collection. With Ollama, nothing leaves the machine at all.

License and author

MIT. Built by Dipak Kumar, sponsored by Boswas Group. Source: github.com/DipakThakur13/COMU.