Ask My Codebase

Ask My Codebase is a small chat app that runs on your own PC and answers questions about your source code. Type a question in plain English; it runs Claude Code against your repo folders, and you get the answer back with the exact file:line citations.

“In the Payslip, do we show YTD from the previous employer or only the current one?”

Instead of grepping through a dozen repositories by hand, you ask once.

Three things make it different

• No API key, no per-token bill. It shells out to the same claude CLI you already use, authenticated with your existing Claude Code login.
• Read-only by design. The bot may only Read, Grep, Glob, and run git log / git show. It answers questions — it never edits or commits code.
• It runs where your code lives. Your machine is the server, so it can see private repos, databases, and logs without any of it leaving your network.

What you can do

• Ask across all repos or scope to one — a checkbox dropdown picks the repositories; tick none to search everything, tick one for the fastest answer.
• Follow up naturally — the server resumes Claude Code’s session, so “and where is that saved?” just works. New chat resets the context.
• Ask from Microsoft Teams — a Teams bot with its own read-only allowlist, so it can never edit or push.
• Query the database, read-only — production and staging are SELECT-only, rows are capped, and PII columns (salary, NRIC, bank) are masked on every read. The bot can only propose write SQL for a human to approve.
• Read live logs, Jira, Monday.com, and feature flags — answer “any errors for this correlation id?” or “is this flag on in staging?” from the same chat box.
• Flag a wrong answer — 👍 / 👎 on every reply. A 👎 files the Q&A into a review queue for the owner. Nothing is auto-trusted and nothing is auto-saved.
• Run specialist agents — triage, code health, QA scenarios, unit tests, weekly-changes summary, and a risk forecast.
• Get a weekly team briefing — an AI-narrated summary of what the team shipped, on a shareable link.
• Search the web when the repo isn’t the answer — current-events questions get a real web search with a cited source and date, not a stale guess.

Configure without editing code

Everything lives in owner-editable settings (⚙ Settings → System), so the same app works for any codebase: workspace repo root, branding, port, model, tool allowlists, and every AI instruction prompt the bot prepends. Most changes apply live; repo root and port need one restart.

Pointing it at a different codebase is three steps — log in as the system owner, set the repo root, restart once.

Token savers

A per-question model dropdown (Haiku / Sonnet / Opus), a “limit exploration” toggle that caps --max-turns, and an answer cache that serves a repeated question with the same model and repo scope instantly, at zero Claude usage.

Tech

Node.js 22+ using built-in modules only — node:sqlite for users, sessions and history, http, and child_process to drive the Claude Code CLI. Framework-free HTML UI. Playwright powers the live-browser agent; sqlcmd backs the guardrailed SQL bridge. Earlier Python and Node versions are kept in the repo as fallbacks.