v0.0.1 — First public release

First public release of Browser Agent — an open-source runtime for browser agents.

This isn't a chatbot with browser access. It's the harness: an agent loop with layered
context assembly, a typed tool registry, a page-action engine (page_act) that grounds
every action against a fresh whole-page snapshot and verifies the result on the live DOM,
a guardrail layer for permissions and risk tiers, and a skills system for turning verified
runs into reusable, replayable workflows.

Highlights

• Kernel — agent loop as a state machine, per-turn tool budget, abortable at every await.
• Context — base system prompt + fresh per-turn environment + persisted session history.
• Tools — one file per tool, spec derived from its Zod schema, plus an MCP mount (Streamable HTTP).
• Engine (page_act) — whole-page semantic perception (shadow DOM, same-origin iframes,
  occlusion), fingerprint grounding across re-renders, DOM/CDP execution, post-condition
  verification, diagnosis-driven healing.
• Observed-facts reconciliation — pre/post DOM diffs overrule model predictions in both
  directions, so a "success" that changed nothing on the page doesn't count.
• Guardrails — risk tiers, per-tool authorization memory, site allow/blocklists,
  confirmation on dangerous actions, sensitive-value redaction, full audit log.
• Skills & batch — verified runs become parameterized, deterministically replayable skills
  with per-row page verification.
• Model port — OpenAI-compatible / OpenAI Responses / Anthropic providers, native tool
  streaming with a prompted-JSON fallback, and a deterministic scripted mock for tests.

See CHANGELOG.md and
docs/architecture.md
for details.

Install

Download browser-agent-0.0.1-chrome.zip below, unzip it, then load it as an unpacked
extension via chrome://extensions → Developer mode → Load unpacked.

Requirements

Chrome (or any Chromium-based browser with MV3 support). Bring your own model API key —
nothing is sent anywhere except the model provider you configure.