Browser Harness

Quick Verdict

Browser Harness is best for developers who already use coding agents such as Claude Code or Codex and want those agents to control Chrome, write missing helper code, retain site-specific knowledge, and verify browser tasks. Its design is intentionally thin: one Chrome DevTools Protocol connection plus an editable agent workspace, rather than a large library of fixed browser actions. That gives a coding agent considerable freedom, but it also leaves more responsibility for selectors, assertions, permissions, and recovery with the operator.

The repository was updated on July 20, 2026, and stable release v0.1.6 shipped on July 17 with browser-task recordings. However, the project was only created in April 2026 and remains a fast-moving 0.x tool. Teams should treat its APIs, connection flow, and domain-skill conventions as early-stage and regression-test upgrades. The local harness is MIT-licensed and free. Browser Use Cloud is an optional paid service for isolated, headless, or concurrent browsers rather than part of the open-source license.

Best For

• Engineers who want a coding agent to operate admin panels, forms, browser tests, or repetitive web flows.
• Browser-agent developers who prefer direct CDP access over a heavily abstracted automation framework.
• QA teams that want screenshots, page-state checks, and recordings alongside an agent-driven test.
• Teams prepared to maintain deterministic assertions and recover when sites, selectors, or authentication flows change.
• It is not ideal for no-code users, buyers requiring a mature enterprise SLA, or anyone planning to expose a personal browser profile and all of its sessions to an autonomous agent.

Key Features

• Thin CDP connection: The harness attaches to Chrome or Chromium through a WebSocket and leaves low-level page capabilities available to agent-written code.
• Editable helper workspace: When a task needs a missing operation, the agent can add a helper to agent_helpers.py and reuse it later.
• Domain skills: Optional domain-skills retain site-specific flows, selectors, and edge cases so future runs do not rediscover everything.
• Local or cloud browsers: It can attach to a running local Chrome, a separately launched debug browser through BU_CDP_URL, or a Browser Use Cloud session.
• Task recordings: Release v0.1.6 can record actual browser frames, clicks, cursor movement, and typing, then condense a long run into a reviewable video. Local recording remains off by default.
• Coding-agent setup: The official setup prompt is designed for Claude Code or Codex and installs the tool with uv on Python 3.12.
• Agent-generated adaptation: The system can write missing browser code during execution rather than requiring every action to be predefined.

Use Cases

Browser Harness can drive end-to-end smoke tests, admin-console checks, form completion, file uploads, browser regressions, and recorded product demonstrations. A reliable test should not accept the agent’s statement that it succeeded. Read the DOM or page state, assert the final URL and key values, check the backend outcome when possible, and keep a screenshot or recording. Video improves explainability, but it is not a complete audit trail because network requests, background scripts, extension behavior, and hidden side effects may not be visible.

Web agents face direct prompt-injection attacks. Text in pages, emails, tickets, and documents may instruct an agent to reveal credentials or exceed its task. Use a dedicated browser profile and test account, allowlist required domains, disable password managers and personal extensions, and isolate cookies, API keys, and downloads. Page content must never grant itself new permissions. Login, upload, send, delete, order, payment, and final form submission should require explicit human confirmation, especially when a saved domain skill makes the flow easy to repeat.

Pricing

The local Browser Harness is free under the MIT license. Browser Use Cloud is optional. Its public pricing on July 21, 2026 listed Free at $0 with 3 concurrent sessions, Dev at $29 per month with 25, Business at $299 with 200, Scaleup at $999 with 500, and Enterprise by quote. Public usage rates listed cloud browser sessions at $0.02 per hour and proxy bandwidth at $5 per GB. Plans and rates can change, so production estimates should be checked against the live billing page.

Model tokens, bandwidth, captcha handling, recording storage, and failed reruns can cost more than the browser session itself. Set limits for steps, session duration, concurrency, and daily spend. A purchase workflow should also have a maximum transaction amount and an approval immediately before payment. Cloud isolation is useful, but it does not remove the need for credential boundaries, egress restrictions, and post-run cleanup.

Pros

• Thin, editable architecture lets a coding agent adapt to unfamiliar pages without waiting for a framework release.
• Clear paths exist for a local browser, a dedicated debug browser, and hosted cloud browsers.
• Domain skills can preserve real site knowledge and reduce repeated exploration.
• Version 0.1.6 recordings make long tasks easier to inspect, and local recording is opt-in by default.
• MIT licensing keeps local use free and permits modification.
• Recent commits and a recent stable release indicate active maintenance on the review date.

Cons

• The project is young and still on 0.x, so compatibility and operational maturity need continued validation.
• Direct control of a real browser is highly privileged; existing cookies and accounts can be misused or exposed.
• “Self-healing” means agent-written adaptation, not deterministic correctness or permanent selector stability.
• Isolated cloud browsers and managed captcha capabilities are separate services with additional cost and data boundaries.
• Recordings do not capture every network or backend side effect and may themselves contain sensitive information.
• Site changes, A/B tests, anti-automation systems, and model variability can invalidate saved domain skills.

Alternatives

FAQ

Is Browser Harness free?

The local harness is free under the MIT license. Browser Use Cloud is optional and has a free tier plus paid plans and usage charges for browser time and bandwidth.

How is it different from Browser Use?

Browser Harness is a thin, editable CDP layer that expects a coding agent to add helpers. Browser Use is the broader browser-agent framework and managed cloud product.

Can it connect to my everyday Chrome profile?

It can, but high-risk tasks should use a separate profile, test account, and isolated download directory with personal extensions and password managers removed.

Are task recordings a complete audit log?

No. Recordings show visible interactions. Also retain the run ID, prompt version, domains, important tool calls, assertions, and backend events, and apply access and retention controls to the video.

How can I reduce browser prompt-injection risk?

Treat page content as untrusted data. Use a domain allowlist, minimum tool permissions, and temporary credentials. Reject page instructions that request secrets, rule changes, data transfer, or access to a new domain.

Does self-healing mean a workflow will not break?

No. The agent can attempt to write a fix and preserve what worked. Site changes, captchas, permissions, and model variance still require deterministic tests, retries, and human takeover.

Bottom Line

Browser Harness is attractive because it gives browser control to an agent that can write code and lets helpers and domain skills evolve through actual work. It is useful for experiments, testing, and bounded automation, but it remains a young 0.x project. Pilot one complete flow with a dedicated account and domain allowlist, then inspect its assertions, recording, recovery, credential isolation, and cost caps. It should reach real business systems only after submission, deletion, purchase, and payment actions reliably stop for human confirmation.