An AI-native browser execution runtime. It sits between AI agents and your real Chrome or Edge, and makes logged-in web apps programmable, safe and cheap to automate. Repeated tasks get faster each time they run.
Status: implementation in progress (P1-P3 merged into main). Start at
docs/README.md.
first run: agent → bos click "the April invoice" → observe → small LLM picks element → act → learn
later runs: bos task run download-invoice --param month=May → cached locators → validate → act (0 LLM calls)
- Real browser, real logins: persistent Browser-OS profiles in Chrome/Edge. You log in once by hand.
- Cheapest mechanism first: learned locator → deterministic match → small LLM → human. Vision comes later.
- Compact pages: the accessibility tree and layout become about 1k tokens of element lines, not megabytes of HTML or screenshots.
- Learns and self-heals: tasks are stored as parameterized trajectories, validated on every replay and repaired when a site changes.
- Safe by construction: secrets are never stored or sent to models; risky actions need confirmation; CAPTCHA and MFA always go to a human; it never hides being automation.
Read AGENTS.md, then docs/CODING_AGENT.md, then pick a task from docs/tasks/README.md.
| Product | docs/PRD.md |
| Architecture | docs/ARCHITECTURE.md, docs/specs/ |
| Decisions | docs/adr/ |
| OSS strategy | docs/OSS_STRATEGY.md |
| Plan | docs/ROADMAP.md, docs/tasks/ |
| Quality | docs/SECURITY.md, docs/PERFORMANCE.md, docs/BENCHMARKS.md, docs/TESTING.md, docs/COMPATIBILITY.md |
| Origins | docs/history/ (non-authoritative brainstorming) |