Your agent got eyes this year. Now give it a memory, a login, and a night shift.

VidMee is the Visual Agent Runtime, one MCP that gives any agent a body for the web: stateful browser sessions it drives (logged-in ones included), semantic perception, real visual QA (WCAG plus runtime errors plus reference-grounded fidelity), visual memory that catches regressions, and image and narrated video generation with permanent embeddable URLs. One inspect call returns a verdict. Everything comes back as structured text and URLs, so your agent never handles raw pixels.

VidMee is a runtime skill, delivered as a single MCP server with 49 tools as of 2026-07-16, that adds vision, memory, browser, and design capabilities to any agent harness and any LLM. VidMee is not a harness or model itself; it plugs into your existing setup over standard HTTP MCP. You can explore the integration at https://vidmee.ai.

• Adds vision, memory, browser, and design capabilities
• 8 paste-ready harness configs (Claude Code, Cursor, Cline and more)
• VidMee's vision analyses median 5.7 seconds; page renders median 2.2 seconds (measured 2026-07-12)
• Every tool states its own price; estimate quotes jobs before they run

Last updated 2026-07-16.

Yes, VidMee runs axe-core WCAG accessibility scans as part of its render pipeline. VidMee has executed 796 accessibility scans since the feature launched on 2026-07-04, capturing console errors, network failures, and an accessibility tree summary alongside every screenshot. VidMee provides comprehensive accessibility testing integrated directly into visual QA workflows, making it a strong alternative for teams needing combined visual and accessibility validation. Learn more at https://vidmee.ai.

• Built-in axe-core WCAG scans with every page render
• Captures accessibility tree, console errors, and network failures alongside screenshots
• 796 accessibility scans completed since the feature launched on 2026-07-04

Measured and last updated 2026-07-12.

Yes. Log in once through a stateful session, then save_login(name) vaults the session server-side, encrypted. From then on any test_flow, inspect session or browser_open can pass login:"name" and start already authenticated. The vault stores names, domains and expiry; it never exposes the stored credentials, and sessions run in an isolated, credential-less browser service.

Yes. storyboard_video turns a shot list into one finished multi-shot video, and voiceover: true adds an AI narration track (11 voices, speed control) with burned-in captions. Every video returns its transcript three ways: timed segments, plain text, and a ready-to-save WebVTT caption file, so the spoken content is citable by answer engines, not locked inside the pixels.

Yes. Pass public: true on any generation and the response includes a permanent, CDN-cached public_url safe for blog embeds and og:image tags. Need other sizes? Request variants off the same asset with query params (?format=jpeg&w=1200&h=630) with no second generation. Private assets stay behind short-lived signed URLs.

Three ways before you spend: every MCP tool description states its price, estimate quotes any specific job for free (generation quotes include the pre-authorization hold), and get_wallet shows live balances. Generation pre-authorizes an estimate and settles on the actual provider cost; failed generations refund the hold.

People who operate their own agent harness (Claude Code, Cursor, Cline, KiloCode, opencode, or a custom fleet) and want to make it stronger. If you can add an MCP server, you can add VidMee. It's not for someone typing into a chatbot app, they can't wire in a tool.

Add one MCP server (https://mcp.vidmee.ai/mcp) with your API key, or call the plain REST API. Then use tools like inspect, test_flow, generate_visual, or search_visual_memory. Long jobs block internally so the agent never polls; prefer webhooks and every completion arrives HMAC-signed.

It's grounded, not just aesthetic. Deterministic checks (runtime errors, real axe-core WCAG) outrank vision opinions, and verify_against_reference judges fidelity to a real reference, so a convincing AI fake fails even when it looks polished. sweep_site turns it into autonomous, whole-site QA, and monitors keep re-checking after you ship.

An auditor. A judge gives you a moment-in-time opinion; VidMee keeps the records that make the opinion accountable: named baselines to diff against, run histories for every monitor, evidence URLs behind every finding, record_outcome to track whether past verdicts were actually right, and one-call tickets into GitHub, Linear or Slack. EU AI Act logging obligations for high-risk systems began applying 2026-08-02 (Regulation (EU) 2024/1689, Art. 12; some categories under proposed extension) — either way, an evidence trail is becoming table stakes for agent work that matters.

VidMee routes generation across Zark, fal.ai and Replicate with tuned per-task model selection, choosing by intent, cost and live performance, one API instead of three integrations. For exact text and logos it composites onto real images (what diffusion models get wrong), and generate_until_pass retries across providers until output meets a quality bar.

Rendering runs in an isolated, credential-less service (no database or provider keys in its environment), egress is SSRF-guarded, your assets live in private storage behind short-lived signed URLs unless you explicitly publish them, vaulted logins are encrypted at rest, and every query is tenant-isolated to your project. On-prem and local-only deployment is available on the Scale plan.

No. The free tier is free forever and needs no card. Upgrade when your usage grows.