Agent Browser Core

**Judgement** A strong, production-leaning skill for AI-driven browser automation when you already have or can install the `agent-browser` CLI. It favors determinism, small snapshots, and structured JSON over ad-hoc scripting, making it a good fit for serious agent workflows rather than quick one-off browsing. **What it does well** - Wraps `agent-browser` into an AI-friendly interface with clear patterns: snapshot → act via refs → re-snapshot. - Emphasizes JSON output and structured commands, which is ideal for OpenClaw-style orchestration. - Built on a Rust-based CLI (with Node.js fallback), so performance should be solid for large or repeated workflows. - Provides substantial documentation (overview, command map, safety, workflows, troubleshooting) that should help reduce trial-and-error. **Key risks / limitations** - Hard dependency on having `agent-browser` CLI and a compatible browser runtime installed and correctly configured. - Not a full SDK: you **cannot** do arbitrary custom JS integrations or deeply customized browser logic from within the skill itself. - Not suitable for heavy media workflows or streaming large uploads; workflows expecting rich media handling may hit limits. - Browser automation is inherently brittle if sites change structure; success depends on good workflow design and selectors. **Safety posture** Safety is relatively strong compared to many browser-automation tools: - Defaults discourage dangerous flags (`eval`, `--allow-file-access`, custom `--executable-path`, arbitrary `--args`). - Recommends blocking localhost/private networks and using domain allowlists. - Calls out credentials/tokens as secrets and warns against unnecessary cookie/storage mutations. Residual risks remain if a calling agent or human explicitly relaxes these safeguards, especially around file access or private network targets. **Best-fit scenarios** Use this skill when: - You need **deterministic**, replayable browser steps for agents (e.g., internal tools, dashboards, authenticated sites). - You want compact snapshots and JSON I/O that integrate cleanly into OpenClaw pipelines. - You are willing to manage the `agent-browser` CLI runtime and follow the documented safety/operational guidance. Less suitable if you want a full programmable SDK, arbitrary JS injection in pages, or rich media streaming workflows.