# Check a web app the agent is building in a real browser

> Check a web app the agent is building in a real browser: 4 companies fully do it, 15 with a limit, 30 carry the tag. Every listing in Browser automation with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/browser.check-web-app/
- Markdown: https://www.anchorterminal.com/niches/browser.check-web-app/index.md (~2,350 tokens)
- Slim: https://www.anchorterminal.com/niches/browser.check-web-app/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/browser.check-web-app/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 4 | 15 | 30 | 22 | 17 |

## What counts as done

The agent reports pass or fail for each named step with a screenshot or snapshot as evidence.

Measured, for our own tests: a 10-step flow takes at most 20 browser actions, and its page reads do not exceed about 50k tokens.

How agents ask for it: open the app and check the signup flow works; run the UI through the browser and tell me what breaks; test this page as a new user; take a look at the page and compare it with the design

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **agent-browser (Vercel Inc.)**: snapshot -i, -c and -d cap the tree, --max-output truncates page text, screenshot writes a path and console reads errors. acc-48 and acc-40 use it for precisely this; acc-72 reports flaky refs on dynamic pages.
- **Browserbase (Browserbase, Inc.)**: The Browse CLI runs a managed local browser (--local) or a remote session, prints accessibility snapshots with cached refs, and the ui-test skill reports passed, failed and skipped tests with failure screenshots under .context/. Remote sessions still cannot reach localhost.
- **Chrome DevTools MCP (Google LLC)**: take_snapshot returns a filtered tree by default (verbose false) or writes to filePath, actions default includeSnapshot to false, take_screenshot writes a file, and a local Chrome reaches localhost. acc-72 puts the tool definitions at about 17k tokens; --slim trims them.
- **Kane CLI (TestMu AI (formerly LambdaTest))**: A plain-English objective runs in local Chrome, so localhost works, and every run seals an evidence pack with per-step screenshots, annotated copies, console and network logs and a pass or fail result card; saved test.md files replay cached steps. Runs spend account credits.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Airtop (Switchboard Visual Technologies, Inc. (dba Airtop))**: A cloud browser driven by page-query, click and type calls over Puppeteer or Playwright, with recordings per window. It cannot reach localhost without a tunnel, and there is no accessibility snapshot or ref tool, so each step check is an AI call billed in credits.
- **Anchor Browser (Anchor Forge Inc.)**: The hosted MCP's anchor_* tools give navigate, snapshot with accessibility refs, click, type, screenshot, console and network logs, and sessions record MP4 by default. The browser is in Anchor's cloud, so a localhost app needs a tunnel, and no snapshot cap or depth option is documented.
- **Browser Use (Browser Use, Inc.)**: The cloud is a goal agent (run_session over MCP, POST /runs) whose browsers cannot reach localhost. The open-source library runs locally and can attach to the person's Chrome, but it too is an agent loop, so per-step pass or fail evidence comes from its history, not a step tool.
- **Browserless (browserless.io, Inc.)**: browserless_agent snapshots interactive and informational elements with maxElements and returns only changes after the first, screenshot toDisk saves a file reference. The browser is in the cloud, so localhost needs a tunnel, and Free sessions end at 2 minutes with 2 concurrent.
- **BxPlaywright (Ortus Solutions, Corp.)**: BoxLang AI agents get browser_visit, browser_snapshot (compact tree with refs, limitable to one selector), browser_click, browser_fill and browser_screenshot against a local browser that reaches localhost. Only BoxLang agents get these tools; others get Microsoft's Playwright MCP, and the in-process snapshot has no depth or size cap.
- **Functionize MCP server (Functionize, Inc.)**: Studio runs the test in Chrome, Firefox or Edge with per-step pass or fail, screenshots and video; the coding agent reads the status rather than driving a browser. A local app needs Functionize's tunnel, a generation is green or yellow never red, and creating a test spends unpublished credits.
- **Hyperbrowser (S2 Labs Inc.)**: A cloud Chromium driven over CDP, WebDriver or computer actions, with rrweb recording on by default and a live view. No snapshot or ref tool of its own (the MCP has run_browser_task and web_fetch), and a localhost app needs a tunnel.
- **Kernel (Kernel Technologies, Inc.)**: Cloud Chromium driven by Playwright, CDP or computer controls with replays recorded by browser id; the MCP server launches browsers and runs Playwright code rather than snapshots. A localhost app needs a tunnel, and acc-36 measured 0.94 s to create a browser.
- **Lightpanda (Selecy SAS)**: A local MCP server with goto, tree, interactiveElements, click, fill, consoleLogs and a screenshot tool reaches localhost cheaply. There is no graphical rendering, iframes, workers and stylesheets are off by default, and acc-64's agent task failed on a link click.
- **Notte (Notte Labs Inc.)**: The MCP gives browser_observe (element IDs), browser_action, browser_scrape and browser_screenshot over a cloud session, and every session records an MP4. The browser is remote, so a localhost app needs the CLI tunnel command, which was not read.
- **Playwright MCP (Microsoft Corporation)**: browser_snapshot takes depth, target and filename, browser_find returns matches instead of a tree, --snapshot-mode none stops action responses carrying one, and screenshots land in --output-dir. acc-01, acc-08, acc-09 and acc-10 use it for UI checks; acc-04 spent over 1M tokens on a 20-step test and acc-03 found it slow.
- **Puppeteer (Google LLC)**: A Node library: the agent writes a script that opens localhost, acts through locators and saves screenshots, so evidence per step is whatever the script prints. No snapshot or ref tool; page.accessibility.snapshot is the compact read. acc-58 let an agent troubleshoot web code this way in Docker.
- **Selenium (Software Freedom Conservancy)**: The agent writes WebDriver scripts in one of five languages and attaches screenshots on failure; there is no accessibility snapshot or ref tool, so page reads are page source or element queries. acc-59 had Claude write JUnit and Selenium tests with no result reported.
- **Skyvern (Ikonomos Inc.)**: An agent run records video, a screenshot and reasoning per step and Browser Address lets Skyvern drive your own browser over CDP, so localhost is reachable. Each action costs 30 credits, acc-54 waited 4 to 5 minutes for a six-field form, and a new account needs a card.
- **Steel (Nen Labs, Inc.)**: The steel CLI drives a session with snapshot refs (@e3), waits for downloads and screenshots, and every session has replay, console and network tabs plus Agent Traces. The browser is in Steel's cloud, so localhost needs a tunnel, and Launch sessions end after 15 minutes.
- **TinyFish (Tiny Fish, Inc.)**: A how-to runs a numbered user flow against a preview deploy or tunnelled localhost in Strict mode and returns passed as JSON, telling you to check passed rather than run status. Strict mode and max_steps are beta-gated, and steps cost $0.016 each.

- not this job: CurlShot (MEDIAPRONET SRL): One capture per call of a public URL; the MCP tool has no click, type or snapshot and cannot reach localhost. The REST click and wait_for_selector parameters stage one render, not a flow.
- not this job: Pomerado (Pomerado AI, Inc.): Builds tools for public consumer sites and portals from a site_url origin; nothing drives a localhost or preview app step by step, and no screenshot or snapshot tool exists.

## What stops agents finishing

- a 20-step test that spends over a million tokens on snapshots (acc-04)
- refs that go stale after navigation so the agent retries blindly (acc-15)
- canvas and custom widgets outside the accessibility tree
- a tool that reports a saved screenshot that was never written (acc-62)

## Around it

Before: agent-harnesses.spec-to-build-loop, agent-harnesses.fix-failing-ci. After: agent-harnesses.spec-to-build-loop, browser.capture-evidence.

## Where this comes from

Weight 22 of 100 in Browser automation: 15 of 61 accounts, the most common job on Hacker News among coding-agent users (acc-01, acc-03, acc-04, acc-08 to acc-11, acc-13, acc-16, acc-37, acc-39, acc-40, acc-48, acc-59, acc-61)

As data: https://www.anchorterminal.com/api/v1/use-cases/browser.check-web-app.json, or the MCP tool get_use_case with id browser.check-web-app. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
