v1.4.0 · Windows · one portable file
GitHub, as a desktop app
A single exe that does the work the website makes you do by hand — with an assistant that refuses anything outside GitHub before the request is sent.
63 MB · no installer · no Python
pages
assets
tools
build_docs.py
verify_search.js
core/github_api.py
ui/pages/repo_admin.py
10 pages, one executable
README Studio
A README written from the real repository — its tree, its languages, its topics — refined on request and committed once you approve.
Repositories
Create, edit, archive, fork and star, with files attached at creation and removable afterwards.
Inside one
Browse the file tree, create branches, read tags, set collaborator permissions, inspect webhooks.
Issues & PRs
Draft an issue, triage a whole queue at once, reply to review comments, close and reopen.
Releases
Raw commits turned into release notes, with Conventional Commit messages and branch names.
Scoping
Off-topic requests are refused before they reach the model, not after the answer comes back.
How it works
Four steps, then you never touch the website again
The token decides what it can reach. Nothing else about your account is touched, and every write is confirmed before it happens.
Connect a token
A fine-grained personal access token. Encrypted with Windows DPAPI the moment you paste it, and readable by nothing but this app on your account.
Point it at a model
Any OpenAI-compatible endpoint — OpenAI, OpenRouter, Groq, Together, Ollama, LM Studio. A local model needs no key and the app works offline.
Ask for the work
Write a README, triage the issue queue, turn forty commits into release notes. The model sees the repository, not a guess about it.
Review, then commit
Every write shows you the file, the branch and the message first. Nothing lands on GitHub that you have not seen.
The assistant's boundary
It will not answer anything that is not about GitHub
Not a preference — an enforcement. The request is classified before it is sent, so a question about recipes or share prices never costs you a token and never comes back with a confident wrong answer.
Before the request
The prompt is classified on your machine. Off-topic and it is dropped there.
In the system prompt
The model is told the boundary in the same words every time, so it does not drift.
On the answer
What comes back is checked too. A refusal is shown as a refusal, not as a blank.
What it will not do
Said plainly, so you are not looking for it
The scope limit is a feature. These limits are the same kind of thing — stated rather than discovered.
Scope enforced before sending Runs offline with a local model No telemetry
It will not manage SSH keys, create OAuth apps, or edit webhooks — a hook you added here would start firing at a server you do not control.