Describe a change in plain language. The extension makes the edits, shows you a diff, and — once you approve it — commits to a new branch, pushes, and opens a pull request on GitHub.
One flow, from the chat panel or the LLM PR Assistant: Generate PR from Prompt command:
Plan — larger tasks are broken into a multi-step plan first.
Edit — Claude edits your files itself through a capped tool-use loop (read_file, write_file, list_files), scoped to the repository root. There are no shell or network tools.
Review — changed files open in VS Code's diff view. Nothing is committed until you pick Apply & Create PR.
Ship — approved changes go onto a new branch, are committed and pushed, and a PR is opened against your base branch.
What it is not
This is not a full in-editor coding agent. There is no autocomplete, no always-on assistant editing as you type, no MCP integrations, and no browser control. It does one thing: turn a prompt into a pull request you review on GitHub.
Requirements
VS Code or Cursor ^1.85.0
An Anthropic API key — kept in VS Code's SecretStorage, not in settings
A GitHub account — sign in from the extension; the token is kept in SecretStorage too
A git repository with no uncommitted changes (changes under .vscode/ are ignored)
Features
Guided onboarding in the editor for the Anthropic key and GitHub sign-in.
Repo and base-branch auto-detection, overridable in settings.
Streaming progress — long calls report into the chat panel and output channel instead of going quiet for a minute.
Structured planning — the plan call uses a JSON schema, so a bad plan fails as a schema error rather than a parsing guess.
Prompt caching for repeat calls against the same repo; cache reads and writes are logged to the output channel.
Relevance-ranked context — tracked files are scored against your prompt and the top 12 are included (up to 3,000 characters each), on top of the file list, README.md, and package.json.
Diff preview and approval gate before any commit or push.
Model picker — LLM PR Assistant: Select Claude Model lists models from the Anthropic API rather than a hardcoded set.
Adjustable reasoning effort via llmPrAssistant.effort (default high).
Distinct error messages — a model refusal, a repository that cannot be read, and a repository with no tracked files each say so specifically.
Docs
Usage, commands, and settings: docs/usage.md
Changelog: CHANGELOG.md
Contributing and local development: CONTRIBUTING.md
