AnimaWatch MCP server
Commit Nov 13, 2026Detects web animation jank, stuttering, and visual artifacts using AI vision and Playwright browser recording.
Also known as rishitank/animawatch
- Stars
- 0
- Tools
- 11
- Downloads, 30 days
- β
- Latest version
- β
- Last commit
- Nov 13, 2026
- License
- MIT
What is AnimaWatch MCP?
AnimaWatch detects web animation jank, stuttering, and visual artifacts using AI vision and Playwright browser recording. It exposes 11 tools, such as analyze_fps, watch, watch_with_device.
Summary generated from the project's README
About AnimaWatch MCP server
Detecting web animation jank, stuttering, and visual artifacts from inside an AI client is what animawatch enables for frontend developers and testers. It uses Playwright to record browser interactions as video and then employs vision AI to inspect the footage for timing issues or visual glitches. Users can choose between using the Gemini 2.0 Flash free tier via an API key or running the process entirely locally using Ollama with models like qwen2.5-vl:7b.
Developers interact with the server through specific tools like watch, screenshot, and check_accessibility. The system can record interactions, analyze existing video files, or perform static page analysis. It also provides access to stored recordings and analysis results through dedicated MCP resources. This setup allows for automated visual testing of complex UI behaviors, such as modal transitions or scroll smoothness, by mimicking how a human tester would observe a screen.
What you can do with it
- 1Diagnose animation stutteringAsk the AI to watch a specific UI transition to find exact timestamps where jank occurs.
- 2Verify interactive UI elementsInstruct the client to click a menu and monitor the resulting slide-in animation for smoothness.
- 3Perform visual accessibility auditsRequest a check on a webpage to evaluate contrast and readability through visual analysis.
- 4Analyze existing video recordingsSubmit a previously recorded video file to the server for automated visual inspection.
Written from the project's README. Every use case is checked against a quote from it.
AnimaWatch MCP tools 11
Detected from the READMEanalyze_fpswatchwatch_with_devicescreenshotget_performance_metricscompare_screenshotslist_devicescheck_accessibilityanalyze_with_consensus_toolanalyze_videorecord
| analyze_fps | β | |
| watch | β | |
| watch_with_device | β | |
| screenshot | β | |
| get_performance_metrics | β | |
| compare_screenshots | β | |
| list_devices | β | |
| check_accessibility | β | |
| analyze_with_consensus_tool | β | |
| analyze_video | β | |
| record | β |
Configuration
GEMINI_API_KEYOLLAMA_MODELVISION_PROVIDER
| Name | Description | |
|---|---|---|
| GEMINI_API_KEY | β | |
| OLLAMA_MODEL | β | |
| VISION_PROVIDER | β |
Automated checks
Ran Nov 14, 2026Full report βNot verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work Β· What to fix for more points
- Source repository linked and reachablerishitank/animawatch is public and reachablegithub.com/rishitank/animawatch βPass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html βPass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/rishitank/animawatch βPass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model11 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNo tool named like delete, drop or execPass
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHub㪠AnimaWatch
An MCP server that checks web animations the way a tester would. It measures dropped frames inside the page, and it can look at a recording with AI vision to find visual glitches.
AnimaWatch is built with FastMCP from the official MCP Python SDK.
β¨ What It Does
YOU: "Check the menu animation on this page for jank"
β
ANIMAWATCH:
1. Opens the page in Chromium and measures every frame in the page
(requestAnimationFrame + Long Animation Frames), against the
display's real refresh rate
2. Optionally records the interaction and asks a vision model
(Gemini with your own key, or Ollama locally) what looks wrong
β
RESULT: "18.4% of frames dropped. A 104 ms long animation frame at 1.3s
came from blockMainThread (app.js). The menu text overlaps the logo
while it slides in."
How it measures jank
Frame timing comes from inside the page, not from the video:
requestAnimationFrametimestamps. A gap of n refresh intervals between two frames means n β 1 dropped frames. The jank percentage is the share of the display's refreshes that got no new frame.- The refresh interval is measured. Before each page loads, AnimaWatch samples the blank tab in the same browser, so it judges a 120 Hz display at 120 Hz. You can also set the rate yourself.
- Long Animation Frames API (Chromium 123+). It lists frames over 50 ms and the script, and where possible the function, that held each one up.
- Compositor frames (optional,
trace_compositor). Chromium's own per-frame verdicts come from a DevTools trace. A CSStransformoropacityanimation runs on the compositor and can stay smooth while the main thread is blocked. The trace shows the difference. - Page load is not counted. Measurement starts once the network is idle.
The video is used only for the vision pass. Playwright records it at a fixed 25 fps, so frame gaps in the video say nothing about the page's frame rate. Versions up to 0.3.1 measured jank from the video, and a perfectly smooth animation scored 98.8% jank there. That method has been removed.
π Features
| Feature | Description |
|---|---|
| π― Frame timing | Dropped frames, frame-time percentiles and long animation frames, measured in the page. No AI or API key needed. |
| ποΈ AI vision | Gemini (your own API key) or Ollama (local) reviews recordings and screenshots for visual problems |
| π± Device emulation | iPhone, Pixel, iPad and desktop profiles |
| π Core Web Vitals | LCP, FCP, CLS (largest session window) and TTFB |
| π Visual diff | Pixel comparison of two pages |
| βΏ Accessibility checks | Visual accessibility review (contrast, readability); needs vision |
| πΎ Resources | Recordings and analyses as MCP resources |
| π Prompts | Prompt templates for animation, page and accessibility reviews |
π οΈ MCP Capabilities
Tools
| Tool | Description | Needs vision |
|---|---|---|
analyze_fps |
π― Dropped frames and long animation frames for a URL, after optional actions | No |
watch |
π¬ Frame timing, plus a vision review of the recording | Optional |
watch_with_device |
π± watch with device emulation |
Optional |
screenshot |
πΈ Screenshot, with a vision review when vision is on | Optional |
get_performance_metrics |
π Core Web Vitals and page-load numbers | No |
compare_screenshots |
π Pixel diff between two URLs | No |
list_devices |
π± Device profiles | No |
check_accessibility |
βΏ Visual accessibility review | Yes |
analyze_with_consensus_tool |
π€ Gemini and Ollama review the same screenshot, and the findings they agree on are reported (not offered when hosted) | Yes (both) |
analyze_video |
π₯ Vision review of a local video file (not offered when hosted) | Yes |
record |
βΊοΈ Record without analysis, saving to a local folder (not offered when hosted) | No |
Actions (click, hover, type, scroll and wait) are passed as a list, for example [{"type": "click", "selector": "#menu"}]. They are validated before any page opens: an unknown action or a missing selector is an error, not a skipped step. Only http and https URLs are opened, plus file:// on a local install. Each wait is capped at MAX_RECORDING_DURATION seconds.
Resources
| URI | Description |
|---|---|
animawatch://recordings/{id} |
Stored video recordings |
animawatch://analyses/{id} |
Stored analysis results |
animawatch://config |
Current server configuration (never the API key) |
Prompts
| Prompt | Description |
|---|---|
animation_diagnosis |
Animation review template |
page_analysis |
Static page review template |
accessibility_check |
Accessibility review template |
π Quick Start
1. Install
git clone https://github.com/rishitank/animawatch.git
cd animawatch
uv sync
uv run playwright install chromium
2. Choose a vision option (optional)
Frame timing, Core Web Vitals, screenshots and diffs work without any of these.
- Gemini, with your own API key. Create a key in Google AI Studio and set
GEMINI_API_KEY. See Cost before you rely on the free tier. - Ollama, on your own machine. See Local vision with Ollama.
- None. Set
VISION_PROVIDER=none. AnimaWatch also runs without vision when Gemini has no key.
3. Configure
cp .env.example .env
# Edit .env if you use Gemini or Ollama
4. Add to Your MCP Client
Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"animawatch": {
"command": "uv",
"args": ["--directory", "/path/to/animawatch", "run", "animawatch"],
"env": {
"GEMINI_API_KEY": "your-api-key-here"
}
}
}
}
Other MCP clients take the same command and arguments. To serve over streamable HTTP instead of stdio, run uv run animawatch --http. It listens on SERVER_HOST:SERVER_PORT (default 127.0.0.1:8765) at /mcp. It accepts only localhost, the bind address and the names in ALLOWED_HOSTS in Host and Origin headers, which protects it against DNS rebinding. The HTTP transport has no authentication, so do not expose it beyond your machine without a proxy that adds it.
π Usage Examples
"Measure the frame rate of https://example.com while the carousel runs"
"Click #hamburger on https://example.com and watch the slide-in animation"
"Watch https://example.com on an iPhone 15 Pro with focus on scroll behaviour"
"Check accessibility on https://example.com"
"Show me the analysis from animawatch://analyses/abc123"
π§ Configuration
| Environment Variable | Default | Description |
|---|---|---|
VISION_PROVIDER |
gemini |
gemini, ollama or none |
GEMINI_API_KEY |
(none) | Your own Gemini API key. Without it, vision is off. |
VISION_MODEL |
gemini-3.8-flash |
The Gemini model. It must accept video. |
OLLAMA_HOST |
http://localhost:11434 |
Ollama server |
OLLAMA_MODEL |
qwen2.5-vl:7b |
The Ollama vision model |
REFRESH_RATE_HZ |
measured | The refresh rate to judge frames against |
BROWSER_HEADLESS |
true |
Run the browser headless |
BROWSER_EXECUTABLE_PATH |
(Playwright's) | Launch this Chromium instead of Playwright's bundled one |
ALLOW_FILE_URLS |
true |
Allow file:// URLs |
VIDEO_WIDTH / VIDEO_HEIGHT |
1280 / 720 |
Viewport and recording size |
MAX_RECORDING_DURATION |
30 |
Upper bound in seconds on every wait |
SERVER_HOST / SERVER_PORT |
127.0.0.1 / 8765 |
Bind address for --http |
ALLOWED_HOSTS |
(none) | Extra host names, comma-separated, that --http accepts in Host and Origin headers, for example a reverse proxy's domain |
The default model was checked on 13 Nov 2026. Google retired gemini-2.0-flash, the previous default, on 1 Jun 2026, and lists gemini-3.8-flash as stable with video input (models, deprecations).
π° Cost
AnimaWatch is open source (MIT), and frame timing needs no paid service. The vision pass is the only part that can cost money:
| Option | What you pay |
|---|---|
| No vision | Nothing beyond your own machine |
| Ollama | Nothing to a provider. It runs on your hardware, and a 7B vision model needs a capable machine. |
| Gemini | Google bills your key according to Gemini API pricing. Video input uses far more tokens than a screenshot. |
About Gemini's free tier. Google offers free quota with rate limits. However, the Gemini API Additional Terms (last updated 28 April 2026) say: "You may use only Paid Services when making API Clients available to users in the European Economic Area, Switzerland, or the United Kingdom." So if you make an app built on AnimaWatch available to users in the UK, the EEA or Switzerland, use a paid (billing-enabled) Gemini key, or use Ollama or no vision. This is a summary, not legal advice.
π Local vision with Ollama
# Install and start Ollama (https://ollama.com)
ollama serve
ollama pull qwen2.5-vl:7b
# AnimaWatch needs the ollama extra, and ffmpeg for video
uv sync --extra ollama
export VISION_PROVIDER=ollama
export OLLAMA_MODEL=qwen2.5-vl:7b
Ollama's vision models take images, not video. For watch and analyze_video, AnimaWatch therefore takes 8 frames spread across the whole recording and reviews each one. This needs ffmpeg and ffprobe on your PATH. The model sees still frames, not motion. Frame timing is measured in the page either way.
βοΈ Apify Actor (single-tenant Standby)
The repository includes an Apify Actor definition. Nothing is published yet.
.actor/actor.jsonturns on Standby mode, with the MCP server atwebServerMcpPath: /mcp.- The
Dockerfileis based onapify/actor-python-playwrightand keeps that image's Playwright version, so the bundled browsers match. .actor/input_schema.jsonholds the user's own Gemini key, as a secret field. Without a key, vision is off..actor/pay_per_event.jsondefines one event,page-analysed, charged once for each page a tool loads and analyses. $0.05 is a proposed price, not a final one.- It is designed to be single-tenant. Each user's server runs in that user's own Apify account, so the user pays the platform compute plus the events, and nothing is shared between users. Hosted, it refuses
file://URLs, does not offeranalyze_video,recordoranalyze_with_consensus_tool(there is no Ollama), and writes no diff images. With vision turned off, any key in the input is dropped. - Known limitation: the hosted browser will open any
http(s)address, including private or loopback ones. Single-tenancy limits that to the caller's own run. A multi-tenant mode would first need request interception that blocks private addresses on every request, redirects included.
Run it locally with uv pip install 'apify>=4.0.0,<5', then uv run animawatch-actor. Set ACTOR_TEST_PAY_PER_EVENT=true to log charges to a local dataset.
ποΈ Architecture
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β AnimaWatch β
β (FastMCP server) β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Lifespan context (one per server, shared by all sessions) β
β βββ BrowserRecorder (Playwright Chromium) β
β β βββ in-page frame collector (rAF + LoAF) β
β β βββ optional CDP trace (compositor frames) β
β β βββ video recording (vision pass only) β
β βββ VisionProvider: Gemini | Ollama | none β
β βββ recordings / analyses β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Tools: analyze_fps, watch, watch_with_device, screenshot, β
β get_performance_metrics, compare_screenshots, ... β
β Resources: animawatch://recordings, analyses, config β
β Prompts: animation_diagnosis, page_analysis, accessibility β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β Transports: stdio | streamable HTTP (/mcp) | Apify Standby β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
π§ͺ Development
uv sync --all-extras
uv run playwright install chromium
uv run pytest # includes tests that drive a real Chromium
uv run ruff check . && uv run mypy src/
The frame-timing regression tests open real pages in Chromium. A smooth CSS animation must score under 2% jank. A page that blocks its main thread for 100 ms every 150 ms must score over 30%. To use a Chromium installed elsewhere, set BROWSER_EXECUTABLE_PATH.
π License
MIT
AnimaWatch MCP questions
What tools does AnimaWatch MCP provide?
It exposes 11 tools, including analyze_fps, watch, watch_with_device, screenshot, get_performance_metrics.
How does AnimaWatch detect animation issues?
It records browser interactions as video using Playwright and sends the video to a Vision AI like Gemini or Ollama to analyze it for jank and stuttering.
Does this tool support accessibility testing?
Yes, it includes tools for visual accessibility analysis such as checking contrast and readability.
What are the costs associated with using Gemini?
Using Gemini via AI Studio is free within certain limits (15 requests per minute and 1M tokens per day).