GitScribe

Turn a GitHub or GitLab repository into an LLM-ready digest, or have a model write a README that maps the whole codebase.

GitScribe turns a repository into something you can paste into a model. Open a GitHub or GitLab repo, pick the files in a tree that prices every path in tokens as you go, and copy the result as Markdown, XML, JSON or plain text. The total updates on every toggle — there is no "calculate" button. Pick the model you are targeting and the panel tells you whether the selection fits its context window, and what to drop first if it does not. It can also do the other thing: have a model read the code and write a README that maps the codebase, so a model answering questions about it later can find things fast. **Two modes, two different privacy stories.** - **Raw digest** — no API key, no cost, and your code never leaves the machine. - **Generated README** — needs a model, costs money per run (shown before you commit), and sends the selected code to the provider you choose. It cannot work any other way. Ollama is the exception: it runs locally and is free. **What else it does** - Digest a pull request or a range (`v1.0...main`) instead of a whole repo. - Split the output into parts that each fit the context window. - Presets and history, per repository, kept in this browser only. - Exclusion rules (lockfiles, tests, fixtures, binaries) that stay visible and struck through, labelled with the rule that caught them — nothing disappears silently. - A live view of your GitHub rate limit, because 60 requests/hour anonymous can go fast on an office network. **What it is honest about** - Token counts are estimates by default, deliberately biased to over-count by roughly 5–8%, because no public tokenizer exists for Claude, Gemini or Grok. With a key it can ask the vendor for an exact count instead. Ollama windows are exact, read from the running server. - Models invent files. Every path a generated README mentions is checked against the files actually read; anything that does not exist is listed under "Unverified references" rather than left to look authoritative. - GitHub caps a comparison at 300 files. When a PR is bigger, GitScribe names the cap and the files it could not read instead of truncating quietly. - It needs the network. There is no durable offline reopen yet. **No backend, no account, no analytics.** Traffic goes to `api.github.com`, `gitlab.com` (or your own GitLab host), Ollama on `127.0.0.1` if you run one, and — only for README generation or exact counts — the model provider you picked. There is nothing of ours in the path to collect anything with. MIT licensed; the source is on GitHub.