Ask questions about any GitHub repo from your browser—without cloning.
You can replace hub with inspect in any GitHub URL to open the corresponding view here.
Site: gitinspect.com
We don’t run a backend that stores your chats or credentials. Session history, model choice, app settings, optional GitHub token, provider keys / OAuth, and usage totals live only in this browser (IndexedDB via Dexie).
The app still uses the network: Your browser calls GitHub’s API directly to load repository data. When you use your own provider keys, chat goes to those providers (you can route via Settings → Proxy). The bundled free tier is different: chat goes through gitinspect’s server proxy (/api/proxy) so we can rate-limit, rotate a shared host key, and reduce abuse—see the next section.
| What it’s for | |
|---|---|
| Settings → Providers | LLM API keys / OAuth. Use this for the models you want and for providers you pay for. |
| Settings → GitHub | Optional PAT stored only in this browser: higher GitHub API rate limits (60/h → 5,000/h) and richer repo metadata where applicable. |
If you have no provider configured, the app still offers a free tier model in the picker (“Free (with limits)”). That path is rate-limited and subject to the host’s terms; traffic goes through gitinspect’s proxy (/api/proxy) so we can enforce limits and run the shared API key on the server—your prompts are not sent straight from the browser to the model host the way they are when you bring your own key.
We use Vercel (hosting) and OneDollar Stats for aggregate traffic and product analytics. These are private to the project—not used to read your chats, prompts, or repository contents. Chat routes are excluded from page analytics where configured.
- Research agent — Pick a repository, chat in natural language; answers are grounded in the code.
- Stack — pi-mono, read-only shell via just-bash, virtual FS from the GitHub API just-github.
- Local first — Agent work runs in a per-tab
DedicatedWorker; durable state stays on the main thread (IndexedDB). - Resilient — Lease ownership, runtime recovery, and interrupted-turn repair on the main thread; the worker improves responsiveness, not hidden-tab guarantees.
- Lazy loading — Nothing prefetched at construction; everything on demand.
- Tree cache — Full repo tree once via Git Trees API;
stat,exists,readdirfrom cache. - Content cache — File contents by blob SHA (content-addressable).
- Smart API selection — Contents API for small files; raw endpoint for large files (>1 MB).
Inspired by Sitegeist, btca & repogrep.
If you want to test paid sync or public sharing locally:
- Create or connect a Dexie Cloud database.
npx dexie-cloud create- or
npx dexie-cloud connect
- Whitelist your local and production origins.
npx dexie-cloud whitelist http://localhost:3001npx dexie-cloud whitelist https://gitinspect.com
- Copy the generated URL, client id, and client secret into
apps/web/.env.examplevalues in your local.env. - Keep
dexie-cloud.jsonanddexie-cloud.keyout of git. They are ignored in.gitignore.
Unauthenticated: 60 requests/hour. With a token: 5,000/hour. Add a token under Settings → GitHub in the app (stored only in your browser). The tree cache keeps real usage low after the first load.
This codebase has been built with a lot of support from AI. Very little is hand-written; GPT 5.4 was used to create this repository alongside that small amount of manual code. (Not proud of code quality rn but it works and will be cleaned up)