AutoCua vs Notte

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.

Notte 70

Graded as a hosted API. status.notte.cc on Better Stack publishes a status feed with component events (20). The feed, read on 2 November, runs from 23 July to 27 October 2026. In the 90 days from 4 August the Scrape API (/scrape) was down for about 32 hours over 14 days, the longest from 20:14 UTC on 23 October to 16:31 UTC on 24 October (20 hours 17 minutes) and from 04:11 UTC on 23 October for 6 hours 5 minutes. CAPTCHA solving and proxies were down for long stretches, and the Session API for 36 minutes in three short outages. On 17 and 28 August the Session API went down with the Scrape API and CAPTCHA solving (and proxies on 17 August) for under 20 minutes, while the page's API reachability check and the Functions, Vaults and Managed Auth APIs read 100 per cent uptime on 2 November. No outage of the API itself or the whole platform appears. Each outage hit one component or a few for part of the service, which this run reads as partial and minor whatever its length, so minor only (20, was 10). The 8 October reading of the page's JSON data file is set aside, since that file is the page's own data endpoint. Concurrency is published per plan (5, 25, 100), and the rate-limit page, re-read on 2 November, lists per-minute, per-hour and per-day windows on session start, agent start and the prompt helpers without the numbers (8). The docs say to retry 429s with bounded exponential backoff and jitter, that 429s carry no Retry-After or X-RateLimit headers while 529 does, and to retry only an operation known to be safe to repeat. The SDK retries selected transient failures, and only the payment endpoint takes an idempotency key. Safe-retry guidance without keys (12, was 10). SLAs are listed for Enterprise without published terms (0). The API answered healthy at version 1.10.3 on 2 November (10). 20 + 20 + 8 + 12 + 0 + 10 = 70.

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.

Notte 74

Graded on the REST API. OpenAPI 3.1.0 at api.notte.cc/openapi.json, version 1.10.3 on 2 November, with 110 paths and 137 operations (25). llms.txt on the site and in the docs, a Markdown twin of every docs page and a keyless documentation MCP server (10). Every operation has a summary, 383 of 1,620 schema properties carry a description, and the guides explain each product (12). 312 schemas with 77 enums and ranges such as max_duration_minutes up to 1,440. chrome_args and extra_http_headers are free-form (11). The errors page shows the body format and its statuses, but 131 of 137 operations document only 422 as an error (8). The spec version matches the release tags. No dated API changelog page was found; changes are visible through GitHub tags, the Node SDK's CHANGELOG.md, whose entries carry no dates, and a dated migration page for the observe deprecation (8). 25 + 10 + 12 + 11 + 8 + 8 = 74.

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.

Notte 75

The browser MCP endpoint has eight compact tools, and scrape accepts selectors and a structured output schema (22). List endpoints take page and page_size up to 100, and sessions filter by only_active (15). Errors are a message and detail with six documented statuses and an x-error-class header on execution failures (11). An idempotency-key header is required on session payments only, stopping a session is documented as safe to repeat, and the MCP docs mark three tools read-only and five as write. We could not list the tools to confirm the annotations (12). Session start has no required fields, and Python and Node SDKs and a CLI are official (15).

AutoCua 60

Free self-hosted package with nothing to buy: 0 + 20 + 20 + 20 = 60.

Notte 40

No x402, MPP or L402 on Notte's endpoints (0). Unit prices are published without login, re-read on 2 November, $0.01 an hour for browser sessions (listed at $0 on 8 October), $0.05 an hour for browser functions, $10 a GB for residential proxies, $0.20 a GB through your own proxy on Startup, and LLM tokens at the model's price (20). Free plan with $10 of credits and no card (20). A person has to sign up in the console to create a key or approve OAuth (0). 0 + 20 + 20 + 0 = 40.

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.

Notte 62

The hosted platform is closed. The core repository is public under SSPL-1.0, which is source-available and not OSI-approved, and the Node SDK is MIT (20). The privacy policy and terms answered with a bot check and were not read. The recordings page says replays are deleted after 24 hours, while an older page in the docs source gives 7 to 30 days by plan. No public DPA found (8). No deprecation policy found. The Node SDK changelog labels breaking changes and the spec marks one field deprecated (5). The trust centre page we read named no subprocessors or data locations, and other regions are an Enterprise option (3).

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.

Notte 56

A Bearer API key from the console, or OAuth on the MCP server with one scope, email. No key scopes or rotation guidance found (18). The browser MCP endpoint leaves out administration tools, three of its tools are read-only, and vaults keep stored credentials away from the model. No read-only key or confirmation step found (9). Pages are untrusted content. The only guidance found is one line in the Claude managed agents guide about limiting where a credential may be sent (4). Session replays, a network log endpoint and usage logs give a per-session trail (11). A valid security.txt, a disclosure policy with safe harbour and a three-business-day acknowledgement, and a SOC 2 Type 2 report on the trust centre. The security page says there is no paid bounty (14).