Your AI agent can't see web pages: the visual runtime you actually need

Updated 2026-10-12. If your agent only reads HTML or DOM dumps, it is blind to layout, screenshots, video, and the actual UI state a human would judge. What you need is not a new LLM and not a new harness: you need a provider-agnostic visual runtime—see, browse, remember, create—that bolts onto the agent you already run. VidMee is that skill layer: one MCP server, standard HTTP, priced per use from $0 with no card required.

Blind agents fail on pixels, not on tokens

Most agent stacks treat the web as text: fetch, parse, maybe a selector. That misses CSS-only states, canvas, screenshots in tickets, recorded flows, and “does this look right after the deploy.” Accessibility and visual regression are pixel problems. A crawler that never executes JS will also miss content; a dedicated crawlability check returns crawler_blind, partial_gap, or mostly_visible so you know whether non-JS indexers even see the page.

The failure mode is familiar: the model invents a selector, clicks the wrong control, or cannot compare tonight’s dashboard to last week’s. Selector memory that caches how UI elements map to selectors for repeat flows is the practical fix for the second half of that—once a human-grade observe step has grounded the control. Visual memory that uses true image-to-image similarity search is the fix for “have we seen this screen before,” not a vector of OCR tokens.

Concrete example: a checkout button that is present in the DOM but covered, off-canvas, or contrast-failed. Text tools report success; a screenshot plus an axe-core WCAG inspect does not. VidMee’s inspect tool runs real axe-core checks, and since the a11y feature launched on 2026-07-04 it has executed 796 WCAG accessibility scans. That is the class of work a language-only agent cannot do no matter how large the context window.

Four capabilities, not another model

You need four capabilities together: see, browse, remember, and create. See means analyzing screenshots, documents, and video—analyze_screenshot is free and takes image_base64 or a public HTTP(S) URL; analyze_visual costs 1 analysis credit (~$0.01) against an asset_id; analyze_document is 5 analysis credits (~$0.05) for a public PDF URL or base64; analyze_video is 10 analysis credits (~$0.10). Browse means a real browser: open a persistent session, act, observe, screenshot, close. Remember means ingesting visuals and searching them by image similarity, plus vaulted sessions you resume by name. Create means generating images and video (Zark, fal.ai, Replicate) with size presets og, X, story, and square, including generate_until_pass until a quality bar is met.

The extract tool pulls any defined schema from any page, which is how you turn a dashboard or design export into structured data without a custom scraper per product. capture_element uses CSS selectors to screenshot a specific element at full resolution and returns an asset_id plus a signed URL that can be fed into composite_image. compare_screenshots takes two asset_ids. That chain—capture, compare, remember—is the mechanism behind visual QA, not a prompt that says “look carefully.”

Trade-off: vision is slower and costed; utility browser calls are free. You should not run a full vision judgment on every mouse move. You screenshot, observe, then spend an analysis credit when a verdict is required.

Bolt it on: MCP, REST, and two-minute setup

VidMee is a Visual Agent Runtime service, not a harness and not an LLM. It is one MCP server with 49 tools (43 tools as of 2026-07-11) behind https://mcp.vidmee.ai/mcp, plus a REST base at https://api.vidmee.ai. It works with any agent harness over standard HTTP MCP. Supported harnesses include Claude Code, Cursor, Cline, and Hermes; eight paste-ready harness configurations are available, and GET /v1/integrations returns them. Stdio-only clients bridge with npx mcp-remote. An llms.txt file lives at https://api.vidmee.ai/llms.txt. Tools include their own documentation and pricing in their descriptions.

Setup takes approximately 2 minutes from API key to first verdict. A free API key is at https://dashboard.vidmee.ai; no credit card is required; pricing starts at $0 with every call priced upfront. Docs: https://vidmee.ai/docs/. The product is dogfooded nightly by the Hermes QA controller through mcp.vidmee.ai. Completion webhooks are HMAC-signed; idempotency keys support safe retries; findings integrate with GitHub, Linear, and Slack.

Latency, measured 2026-07-12: median completed tool call about 9.9 seconds end-to-end; page renders median 2.2 seconds; vision analyses median 5.7 seconds. That is slow enough that you batch judgments, fast enough that an agent loop can wait on a verdict instead of guessing.

Browse like a logged-in human, without storing passwords

browser_open creates a persistent browser session and returns a session_id. It supports passing login= to resume a vaulted authenticated session from save_login/list_logins. browser_save_session exports cookies and localStorage for later resumption; VidMee does not store credentials server-side. Sessions are encrypted and vaulted, resumable by name. browser_act, browser_close, browser_navigate, browser_observe, browser_save_session, and browser_screenshot are free (utility/read). browser_act supports click, fill, hover, select, press, scroll, and wait.

That split matters. Interaction should be cheap; pixels should be billed when you actually judge them. check_crawlability costs 1 analysis credit (~$0.01) and detects content invisible to non-JS crawlers—useful when SEO and agent visibility diverge.

Example flow: browser_open with a named vaulted login, navigate to staging, browser_observe to ground controls, browser_act through a test_flow or sweep_site, screenshot, then analyze or compare against a stored baseline. Selector memory keeps the mapping stable across nights. Assets generated with public:true return permanent, CDN-cached URLs so transcripts stay citable.

Memory, QA loops, and why self-hosting the skill is the point

Visual memory supports image-to-image similarity search via ingest_visual and search_visual_memory. That is how an agent answers “is this the same error toast as last Tuesday” without OCR. Since 2026-07-02, VidMee's automated QA loop has executed 145 QA sweeps across 8 products, judging every captured screen against stored visual baselines. Since 2026-06-28 it has performed 966 vision judgments and 983 page renders. The QA loop runs on VidMee itself: 7 confirmed findings filed against VidMee, 4 already resolved.

That dogfood is the argument for a runtime skill rather than a closed test cloud. You keep Claude, Cursor, Cline, or Hermes; you add see/browse/remember. Competitors in visual testing and hosted browsers exist (Applitools, Percy, Browserbase); they solve adjacent jobs. The distinctive need here is agent-native tools with published transcripts and per-call prices, including free screenshot analysis when you already have the image.

create is optional but closes the loop: generate_visual, storyboard_video with AI voiceover, timed captions and full transcript, generate_until_pass. You do not need generation to fix blindness; you need it when the agent must produce the artifact it just judged.

The bottom line

Give the agent eyes and a browser with memory, not a bigger prompt. Add a visual runtime skill—MCP or REST, 49 tools, four capabilities—so it can screenshot, act, compare, and remember. Start at $0, paste a harness config, vault a session, and require a visual verdict before the agent claims the page is fine.