mrkeyoor.com_
Mon 07 Sept 14:45 UTC
AI Toolsevaluationupdated 07 Sept 2026

AIHawk review

AIHawk is an open-source AI browser agent: you tell it in plain language to browse, click, type and read real web pages, and it drives an actual browser to do it. The problem it solves is slow, brittle, manual web workflows, like checking fares across calendar widgets, or filling forms that lack APIs.

trackingstars / 7d
Verdict

AIHawk installed in 39 seconds, passed all 138 tests in 22 seconds, and exposed zero known vulnerabilities in our sandbox; if you can accept the OpenRouter key and the fact that a browser agent's judgment is only as good as the model you point it at, it is a legitimate shortcut for real web automation.

We ran it

Install✓ · 39s79 packages · 105 MB
Build✓ · 7s
Tests✓ · 22s138 passed · 0 failed · 6 skipped of 138 (pytest)
Known vulns0(pip-audit)
Repo116 files~6,773 lines of source · 2.7 MB · 5 CI workflows · tests dir

Answers from our run

Does AIHawk build from source?

Dependencies installed in 39 seconds (79 packages), and the build succeeded in 7 seconds. We cloned commit 90b5eb5 into a clean Debian container with 3 CPUs and no project-specific setup.

Do AIHawk's tests pass?

Yes: 138 of 138 passed when we ran the project's own test command (pytest). Some failures need services or credentials a bare container does not have.

Does AIHawk have known vulnerabilities in its dependencies?

pip-audit found none in the dependency tree at the time of our run.

Who should not use AIHawk?

Non-technical users who won't touch a terminal, install uv, or manage API keys; this is not a browser extension

What are the alternatives to AIHawk?

browser-use, Playwright, Skyvern. AIHawk installed in 39 seconds, passed all 138 tests in 22 seconds, and exposed zero known vulnerabilities in our sandbox; if you can accept the OpenRouter key and the fact that a browser agent's judgment is only as good as the model you point it at, it is a legitimate shortcut for real web automation.

Setup4/5One-command uvx install after uv setup; our run took 39s.
Docs4/5Wiki with guides, comparisons, and MCP config.
Community4/530,319 stars, fresh release yesterday, 0 issues.
Maturity4/5v0.11.1, all 138 tests passed, MIT, no known CVEs.

Who it’s for

Developers or power users who already run Claude Code, Codex or Gemini CLI and want an MCP server for browser automation
People comfortable with an OpenRouter key who want a standalone chat UI with live browser view
Automation-minded teams that need proxy, persistent profile and seed control for repeatable browser identities

Who it’s NOT for

Non-technical users who won't touch a terminal, install uv, or manage API keys; this is not a browser extension
Anyone who needs deterministic, scripted browser tests rather than an LLM's judgment; Playwright is the right tool
People who cannot accept the ToS risk of automated browsing on sites that prohibit bots
Those requiring an authenticated remote dashboard out of the box; the UI has no auth and changing the host exposes it

Setup reality

Our run looked like a best-case install: on a fresh Debian 12 container with 3 CPUs and 8 GB RAM, clone, uv install, build, and all tests completed in 39 seconds, 7 seconds, and 22 seconds respectively; we measured 79 packages, 105 MB on disk, 138 passed with 6 skipped, and pip-audit found 0 known vulnerabilities. The README's command path is a couple of uvx steps, but real setup means installing uv, running 'uvx invisible-playwright fetch' to get the browser engine, then adding the MCP server or launching the UI with an OpenRouter key. The key handling is intentionally fussy: passing --openrouter-key puts it in shell history, so the README nudges you to a .env or OPENROUTER_API_KEY, and startup lists variable names but never values. On our box, no system packages were missing, but our sandbox was intentionally clean and unprivileged; anyone on a stricter environment may still need to install browser dependencies that the README doesn't enumerate.

AIHawk is a browser agent with a real browser under the hood. You say what you want in plain language, and it browses, clicks, types, and reads the actual web. The problem it solves is not "no API exists", it's that many real workflows live inside calendar widgets, multi-step forms, and login walls where a human's eyes and mouse still matter. The project has two faces: an MCP server that slots into Claude Code, Codex, or Gemini CLI, and a standalone web UI at http://127.0.0.1:8765 where you chat on the left and watch the live browser on the right. Both faces use the same underlying engine, invisible_playwright, which is a Playwright-compatible Python library, plus invisible_core for seed-to-fingerprint consistency.

What happened when we ran it

Our run was on a fresh Debian 12 container with 3 CPUs and 8 GB RAM, cloned at commit 90b5eb5 on 2026-09-06. Install succeeded in 39 seconds, pulling 79 packages and 105 MB on disk. Build succeeded in 7 seconds. The pytest suite completed in 22 seconds with 138 passed, 0 failed, and 6 skipped out of 138 tests. pip-audit reported 0 known vulnerabilities. The repository has 116 files and about 6,773 lines of Python source. Those are all our measured numbers; we did not test live browsing or model output because the sandbox had no secrets and no network access to target sites. The clean run matters, because a browser agent that cannot install and pass its own tests is not worth driving.

Concrete strengths

The MCP story is unusually direct. The README gives one-line commands to add the server to Claude Code, Codex, and Gemini CLI, and the package serves MCP by default with no subcommand. The UI is opt-in, so you are not forced into a local web app if you just want your assistant to browse. Version 0.11.1 landed yesterday and licenses changed: code before 2 September 2026 remains AGPL-3.0, but the current release is MIT, which removes the copyleft objection for commercial use.

Configuration precedence is also thoughtful: --flag beats environment, environment beats .env, and .env beats the default. The startup only reads .env in the current directory, so running from a subfolder cannot silently pick up a different key, and the log names variables without printing values. Passing --openrouter-key puts the key in shell history, but the README tells you to use OPENROUTER_API_KEY or a .env instead. The --profile-dir option keeps logins and cookies across restarts, and --binary refuses to start if the engine binary does not match the pinned seal, so you cannot accidentally run a mismatched Playwright build.

The documentation is extensive without being padded. The wiki has an AI browser-agent landscape comparison, a guide on why agents get blocked, and a piece on what happened to OpenAI Operator. Worked examples and transcripts live in articles/. The invisible_playwright and invisible_core repos are linked separately, so you can use the engine directly if you want code instead of prompts.

Rough edges and honest cautions

The web UI has no authentication out of the box, and the README says as much: changing --host exposes an interface with no auth. Keep it on localhost unless you put an authenticated reverse proxy in front of it. You also need an OpenRouter key, and the default model is z-ai/glm-4.6, a model we could not evaluate from our sandbox; the agent's judgment is only as good as that model, and the tool does not include model evaluation tools.

There is no Dockerfile in the repo. Our install succeeded with uv, but teams that want an image will need to write their own Dockerfile or use uv in a base image. Browser automation on third-party sites carries real ToS and accuracy risk; the README's responsible-use note says to respect rate limits and not submit anything a human has not read, but it cannot enforce that. The repo reports 0 open issues, which is either a well-maintained tracker or a place where community feedback doesn't land. We cannot tell which from a snapshot, but the number itself is zero.

Community health and cadence

AIHawk was created on 2024-08-04 and has 30,319 GitHub stars at review time. The latest release, v0.11.1, was pushed on 2026-09-06T23:48:47Z and the last push was 2026-09-06T23:45:27Z, about a day before this review. There are 5 CI workflow files and a tests directory. Press mentions include Business Insider, TechCrunch, Wired, The Verge, and 404 Media. That active push cadence, combined with a fresh release, suggests the project is not stale, but the 0 open issues leave no public evidence of maintainer response time.

Where it fits in a real stack

AIHawk fits behind your existing assistant or as a local UI, not in front of an API gateway. If you already run Claude Code or Codex, adding the MCP server is a two-command operation and gives your assistant browsing ability without writing a Playwright script. It is best for exploratory, one-off tasks where the page structure is messy and a person's judgment about what to click matters. It is not a replacement for Playwright when you need deterministic, repeatable regression tests, and it is not a replacement for a human approving high-stakes form submissions. Treat it as a trusted pair of hands for low-consequence browser work, and keep the browser headed or the logs visible until you know how it behaves on your sites.

Alternatives

ProjectWhat it isPick it when
browser-use gh↗An open-source Python library for building your own AI browser agents.Pick when you want a code-first library to embed in Python rather than a prebuilt MCP server or UI.
Playwright gh↗The upstream browser automation library with a deterministic API.Pick when you need scripted, repeatable tests without any LLM guessing.
Skyvern gh↗Automates complex web workflows using LLMs and computer vision, with a cloud offering.Pick when you want a hosted workflow engine and visual verification for multi-step forms.

What people are saying

  1. [velocity-scout] feder-cr/AIHawk

Sources

  1. AIHawk GitHub repository
  2. AIHawk wiki

More ai tools reviews

personaplex · feynman · sepia · rf-detr · peft · DeepSpeed · the whole board →