Obra CTO

A local-first MCP server that scores your codebase's Build Readiness. You install it in your own Claude. Your Claude reads your code on your machine and runs your tests there. Nothing is uploaded.

This is the free preview of Obra's role lineup, and it stays free and local. The CTO reads your code and tells you what an investor's technical diligence would find: real security holes, weak design decisions, and the gaps that stop a serious review. The rest of the team (the Obra CFO for funding-ready materials, and more) lives in the Build with Obra community.

The Obra founders' toolkit

Two free, local, open-source MCPs built to work as a pair:

• Obra CTO (this one): is it built? Scores your codebase and surfaces what a technical diligence would flag. GitHub · npm
• Obra CFO: is it fundable? Runs the investment committee on your pitch and turns your verified build into a funding case. GitHub · npm

Run the CTO first. The CFO reads its report as grade-A technical evidence. The method, the deep versions, and Obra in beta live in Build with Obra.

Why local-first

A tool that reads your code should not ship your code somewhere. Your source never leaves your machine. This server makes exactly one kind of network call, and only when you run the dependency check: it sends your package names and versions to the OSV vulnerability database, never a line of your code (you can see the single call in src/deps.ts). Everything else is local: it returns counts, presence flags, a redacted secrets scan, and a score. You can read every line of this server before you run it, which is the point of keeping it open.

What you get

The Obra CTO Score, out of 100, calibrated to your stage (prototype, MVP, or growth), across six dimensions:

Every finding carries an evidence grade: A verified, B multiple sources, C partial or inferred, D claim only, E speculation. The score is built from grade A and C evidence, what is true in your code, not what a deck says. When the CTO runs your tests and they pass, reliability becomes grade A. A deck-scorer can never earn that.

What a run looks like

Point it at a real project and you get a scored report with a ranked risk register. Here is a run on a scrappy Supabase app (anonymized):

# Obra CTO Score: financeapp
## 46 / 100 · Not yet ready
Assessed as: mobile (react-native, expo)  |  Backend: supabase

| Dimension        | Score | Evidence |
|------------------|-------|----------|
| Security         | 6/25  | A |
| Product reality  | 20/20 | A |
| Robustness       | 3/15  | A |
| Architecture     | 9/15  | A |
| Maintainability  | 6/15  | A |
| Deploy readiness | 2/10  | C |

## Top Risks
- [critical] RLS policies defined but never enabled; data may be open to any authenticated user
    Fix: enable row level security on every table, then verify a second user cannot read your rows.
- [high] Access control gap: a privileged action checks only that the user is logged in, not that they own the resource
- [medium] No tests found

On a well-built app it scores high and credits the good engineering. It is calibrated, not a fear machine.

Tools

• scan_project reads the project and returns mechanical signals.
• check_dependencies checks your locked dependencies against the OSV vulnerability database. Only package names and versions leave your machine, never your code.
• run_tests runs your test suite (this executes code, so your Claude asks first) and parses the pass and fail counts.
• prepare_code_review selects your highest-signal files (schema and policy file

... View full README on GitHub