Sweep Ai

Sweep automates the small coding tasks that clog developer backlogs — install the GitHub App, create a GitHub issue starting with "sweep: add unit tests for the auth module", and Sweep reads your codebase, plans the implementation, writes the code, and opens a PR for human review. For agents: programmatically create GitHub issues with the sweep: prefix to trigger code changes (e.g., sweep: fix the bug on line 42 of api.py where null pointer exception occurs). Sweep handles codebase reading, test running, and iterative improvement. Best suited for small-to-medium changes (<200 lines). Open-source and self-hostable. Confidence is docs-derived.

Primary interface: GitHub App (webhook-driven). Trigger: create GitHub issue with title starting with "sweep: {task description}". Sweep automatically: (1) reads repo codebase, (2) writes plan as issue comment, (3) commits code changes, (4) opens PR with changes. @sweep-bot in comment triggers a new attempt. No direct REST API for triggering — agent control is via GitHub API: POST /repos/{owner}/{repo}/issues with title: "sweep: {task}". PR template: Sweep PRs include plan, changes made, test results in PR description. Sweep config (.sweep.yaml): configure included/excluded directories, test command, and assistant behavior. CLI: sweep run "{task}" for local execution without GitHub.

GitHub App auth: install Sweep GitHub App on repository — Sweep authenticates via GitHub App tokens. No Sweep-specific API key. GitHub token needed for creating issues (to trigger Sweep). Sweep Cloud: no API key required — entirely GitHub-driven. Self-hosted: requires OpenAI API key (GPT-4) or Anthropic key. Private repos: GitHub App requires private repo access. Sweep reads full codebase — evaluate data handling for proprietary code before enabling on sensitive repos.

Task success rate: varies by complexity — routine tasks (add tests, fix typo, add field to model) have high success rates; complex refactors may fail or require multiple iterations. Iteration: @sweep-bot in PR comment triggers a new attempt with feedback. Scope: best for <200 line changes; large refactors often fail. Test validation: Sweep runs configured test command and iterates on failures. PR review required: always review Sweep PRs before merging — outputs may have logical errors. Rate limits: Sweep Cloud limits concurrent runs; self-hosted limits by compute. GitHub issue creation latency: Sweep typically responds within 2–5 minutes of issue creation.

docs.sweep.dev covers installation guide, .sweep.yaml configuration, best practices for writing effective Sweep issues, and self-hosting. Getting started: install GitHub App, create issue — first PR in 5 minutes. Sweep Cloud: free tier (5 issues/month on Pro plan). Self-hosted: requires OpenAI API key. GitHub: sweepai/sweep (8k+ stars). Community via Discord (active). Good documentation; best practices guide for writing effective Sweep issues is particularly useful.