AutoCua vs Chrome DevTools MCP

AutoCua 25

Local package. Installs from PyPI as AutoCua with Python 3.10 or later and the supported platforms stated (20). No CI runs the tests; the only workflow, use_case_label.yml, labels issues (0). GitHub issues weren't read, so the open-issue line is scored absent (0). Four tags in a week with no changelog file (5). Version 0.1.9 (0). 20 + 0 + 0 + 5 + 0 = 25.

Chrome DevTools MCP 83

Scored as a local stdio package. Official npm package, 1.12.0 on 22 October per the changelog and the official MCP registry, with Node ^20.19, ^22.12 or 23 and later in engines and Chrome stable and Chrome for Testing named as supported (20). Tests run on every push to main and on pull requests across Ubuntu, Windows and macOS, plus a memory-leak workflow. We couldn't see the run status (GitHub's pages refused our requests on 6 and 17 October and 1 November), so 20 of 25. 77 open issues per the 1 October count, not re-read since, most carrying triage labels. Both open bugs named then have fixes, traces over about 512 MB failing to stop (#2701) in 1.10.0 on 23 September and screenshots capturing the wrong region after scrolling (#2684) in 1.11.0 on 15 October (#2685), and 1.12.0 lists 51 fixes (18). release-please writes the changelog from conventional commits, but 1.8.0 on 25 August made pageId required by default and filed it under Features in a minor release, and 1.11.0 turned the screencast tools on by default, also as a feature (10). 1.0.0 shipped on 18 May 2026 (15). 20+20+18+10+15 = 83.

AutoCua 42

Framework reading. No API reference or typed interface beyond run_agent's documented arguments (5 of 25). No llms.txt (0). The README says what each mode drives and where it runs, and agent_capabilities.md lists what it can do (16 of 20). Modes and device are documented strings (6 of 15). Many examples, no documented errors (10 of 15). Tags without a changelog (5 of 15). 5 + 0 + 16 + 6 + 10 + 5 = 42.

Chrome DevTools MCP 87

Every tool has a Zod input schema with types, required fields and enums (25). Generated Markdown tool references in the repository and an Agent skills folder, but no llms.txt (7). Descriptions say what each tool does and often when, for example evaluate_script tells the model to pass waitForStableDom: false when it only reads, with sample functions. Few say when not to use a tool (16). Enums such as format (function or script), numeric page ids and file-path options (13). Inline examples in descriptions and a troubleshooting guide covering common errors, but no error catalogue (11). CHANGELOG.md for every release since 0.x (15).

AutoCua 42

MIT (30). Says it sends data only to the chosen model provider and keeps everything else locally, and the source agrees (20 of 30). No deprecation policy (0). States no telemetry, and none was found (20). 30 + 20 + 0 + 20 = 70.

Chrome DevTools MCP 82

Apache-2.0 (30). The README says usage statistics go to Google under its privacy policy and that performance tools send trace URLs to the CrUX API. No retention figures for either (20). No deprecation policy. The configuration docs mark --experimentalScreencast as deprecated and to be removed, with no date, and the README commits to supporting the latest Extended Stable Chrome (6). Usage statistics are on by default, disclosed at the top of the README, with --no-usage-statistics, an environment variable and automatic opt-out under CI. CrUX lookups switch off with --no-performance-crux (20).

AutoCua 65

v0.1.9 tagged on 4 November 2026 (30). Four tags in 90 days (20). Issues and pull requests weren't read (0 of 25). The PyPI wheel is current (15). No CI; dependencies pinned per OS in requirements files (0 of 10). 30 + 20 + 0 + 15 + 0 = 65.

Chrome DevTools MCP 93

1.12.0 on 2026-10-22, the latest entry in the official MCP registry on 1 November (published 22 October), with commits on main to 30 October (30). Seven releases in the 90 days from 3 August, 1.7.0 on 10 August to 1.12.0 (20). Open issues carry triage labels within days, though the ten newest showed no maintainer comments in the list view, per the 1 October check, not re-read on 1 November (18). Listed in the official MCP registry as io.github.ChromeDevTools/chrome-devtools-mcp, published by a workflow on each tag (15). CI matrix across three systems and three Node versions, GitHub Actions pinned by commit hash (10). 30+20+18+15+10 = 93.

AutoCua 38

Framework reading (telemetry, approvals, guardrails, sandboxing). Provider keys stored in a local folder with Chrome profiles (10 of 30). The worker tools are an allow-list with no write tool, unseen ids are refused, downloads need a yes, but on Windows the model answers UAC elevation prompts itself through a kernel driver, with no person asked (6 of 20). The scraping ruleset has a prompt-injection instruction to log page text that gives instructions (10 of 15). Runs are logged in the app, with no audit log (7 of 15). No security policy, disclosure route or advisories found (5 of 20, for the vault and certificate handling). 10 + 6 + 10 + 7 + 5 = 38.

Chrome DevTools MCP 63

No credentials to leak and nothing to scope, so the middle band (20). --javascript-evaluation false disables script tools and javascript:, data: and vbscript: URLs, --allowed-url-pattern and --blocked-url-pattern restrict the browser (now enforced across redirects), MCP roots confine file access, and --isolated uses a throwaway profile. 1.11.0 on 15 October released the --file-navigations flag that rejects file: URLs. No confirmation step for writes, and --isolated, header redaction and the file-navigation block are all off by default (14). SECURITY.md says page content comes back as-is and tells users to prefer trusted content or have the client guard against prompt injection, and since 1.11.0 also says the server trusts its working directory, environment and npm configuration, so it shouldn't be started in an untrusted workspace (8). A --log-file debug log only, no per-call audit (3). Reports go through Google's open-source vulnerability reward programme, and the two moderate advisories of 2026 (CVE-2026-53765 and CVE-2026-53766, both fixed in 1.1.0) are still the only ones in OSV for the package on 1 November. 1.12.0 also fixes filesystem paths that parse as http or ws URLs escaping path validation, with no advisory (18).

AutoCua 55

Framework reading. One call with a mode, provider, model and task runs a whole agent (20 of 25). A separate compression agent splices handoff summaries in, and runs are capped and stoppable (15 of 20). Permission and blocked-input errors are reported specifically, with no error reference (8 of 20). No resume or retry semantics described (2 of 20). Sensible defaults (simulators unless device is set), Python only (10 of 15). 20 + 15 + 8 + 2 + 10 = 55.

Chrome DevTools MCP 73

59 tools in the full reference across 11 categories. 1.11.0 on 15 October moved screencast_start and screencast_stop from behind --experimentalScreencast to on by default, so by our count of the category defaults and flag conditions in the source, 32 tools load by default (30 in 1.10.1), unchanged in 1.12.0 on 22 October. Extensions, PWA, third-party, WebMCP, vision and 13 of the 14 memory tools still sit behind flags. More than 30 tools (5, was 15 at about 30). Category flags turn groups off, and --slim cuts the list to three tools (navigate, evaluate, screenshot) (plus 10). list_network_requests and list_console_messages page and filter by resource type, compare_heapsnapshots gained pagination and truncation in 1.11.0, the heap snapshot tools paginate by default since 1.12.0, and large outputs can go to a file path instead of inline (20). Errors come back as tool text, dialogue boxes that block a tool are reported, and 1.11.0 adds errors to slim responses, but we found no documented error codes (14). Every tool sets readOnlyHint (28 true, 39 false in the 1.12.0 source), with no destructiveHint (14). Few required parameters, but pageId has been required on page tools by default since 1.8.0. npm only (10).