Listed from the MCP registry; nobody on Odel publishes it.
Find broken pages in your running web app before users do. Each page bug comes with steps and a fix.
Runs on your machine from its package; the steps are below.

At most: moves money, runs code or commands, deletes data, or sends messages for you.
Odel charges nothing per call. The publisher may charge for its own service.
Runs locally over stdio
This server isn't hosted — your MCP client launches it from a package registry. Use one of the commands below, or drop the config into your client (e.g. Claude Desktop).
Run
npx -y sniff-qa
MCP client config
{
"mcpServers": {
"sniff": {
"command": "npx",
"args": [
"-y",
"sniff-qa"
]
}
}
}English · 简体中文 · 日本語 · Español · Français
Catch broken pages in your web app before users do.
sniff opens your running app in a real browser and clicks like a user.
| Bugs caught before users | Steps to repeat it | Broken merges blocked |
|---|---|---|
Found 21 of 21 planted bugs, 0 false findings: npm run build && npm run proof. | Each page bug comes with its route, the steps to repeat it, and a fix. | --fail-on exits 1 on critical or high bugs, so CI stops the merge. |
Works with: any MCP client, any agent that reads skills, or the command line alone. MCP is how an agent calls a tool.
Tested: Claude Code, Codex CLI. The walk uses a browser, not a model.
npx sniff-qa --url http://localhost:3000
A real report from this repo's test app with planted bugs (2026-11-17, sniff 0.9.3):
VERDICT FAIL: 26 findings (1 critical, 13 high, 11 medium, 1 low), 1 low-confidence finding hidden
Walked 12 pages, checked 13 links, in 55.2s. --all shows the hidden one.
Fix these first:
1. Page returns HTTP 500 on /crash [critical]
2. Page returns HTTP 404 on /dead [high]
3. Page returns HTTP 404 on /missing [high]
Every finding, with the steps that reproduce it:
CRITICAL (1)
• [confirmed] Page returns HTTP 500
/crash (route/broken-page)
- Navigate to /crash
- Server responded with HTTP 500
fix: The route throws server-side. Check the server logs/handler for this path and return a valid page or a proper error page.
shot: sniff-reports/crawl/_crash-desktop.png
Use a dev or staging app: a walk clicks buttons and submits real forms.
What you get · Install · How it works · Works with · Proof · FAQ · Limits
Try it with no install. Start your app, then walk it from a second terminal:
npx sniff-qa --url http://localhost:3000
The npm package is sniff-qa. Do not run npx sniff: that is an unrelated package.
Recommended: the Vercel skills CLI. npx skills is Vercel's open installer for agent skills.
npx skills add Aboudjem/sniff -g
Works in more than 70 agents: pick yours in the menu (supported agents).
In a terminal it shows the agent menu. In a script it installs into every agent folder, so add -a <agent> there.
-g installs for all your projects. Only this project: drop -g.
It installs the 3 skills only. The MCP server is a separate one-line step, shown below.
Or install straight into one agent. Node.js 22 or newer.
Install tested.
npx skills add Aboudjem/sniff -g -a claude-code
MCP server, tested, for all your projects:
claude mcp add -s user sniff-qa -- npx -y sniff-qa --mcp
Only this project: drop -s user.
Install tested.
npx skills add Aboudjem/sniff -g -a codex
MCP server, tested, for all your projects:
codex mcp add sniff-qa -- npx -y sniff-qa --mcp
Install tested.
npx skills add Aboudjem/sniff -g -a cursor
Install tested.
npx skills add Aboudjem/sniff -g -a gemini-cli
Or add the extension, which brings the MCP server too (install tested):
gemini extensions install https://github.com/Aboudjem/sniff
Install tested.
npx skills add Aboudjem/sniff -g -a github-copilot
Install tested.
npx skills add Aboudjem/sniff -g -a opencode
Adds the skills and the MCP server in one step, through the 10x marketplace:
claude plugin marketplace add Aboudjem/10x
claude plugin install sniff@10x
Any other MCP client runs npx -y sniff-qa --mcp. Snippets for each editor are in docs/editors.md.
The first walk downloads about 440 MB once. On a Mac it uses about 730 MB of disk. It gets the browser engine (Playwright, axe-core and Lighthouse) and Chromium, the browser itself. To do it ahead of time, run:
npx sniff-qa install
It ends with the time it took. On a Mac with empty caches it printed 43 s (2026-11-17, sniff 0.9.3). It left 193 MB of engine and 530 MB of Chromium on disk. Sizes vary by platform. To re-measure on macOS:
du -sm ~/.cache/sniff-qa/engine/* ~/Library/Caches/ms-playwright
npx sniff-qa doctor checks Node, the browser and your dev server.
The MCP server starts before those downloads. With an empty npm cache, npx -y sniff-qa@0.9.1 --mcp
answered its first request in 3.1 to 4.9 s (5 runs, 2026-11-17). Version 0.9.0 took 37 to 93 s,
past Claude Code's 30 s connect limit. To re-measure, in a clone of this repo:
npm run cold-start -- 0.9.1 5
A walk prints progress lines, then the report. You can take the result four ways:
| Output | Flag | Use it for |
|---|---|---|
| Terminal report | (default) | Reading the result now |
| HTML page | --report | Sharing: one file with everything inside |
| JSON | --json | Scripts and other tools |
| Exit code | --fail-on | Failing CI on real bugs |
| Exit code | Meaning |
|---|---|
0 | No finding reached the --fail-on severities (default: critical and high) |
1 | A finding reached them, an assert limit was passed, or sniff could not run |
To block bad merges, use --fail-on in CI. Its default is critical and high.
npm install -D sniff-qa adds it to a project. npx sniff-qa ci writes a GitHub Actions workflow
that caches the browser and saves the report. With no app running, its first run only reads your source.
Turn on its Start the app step (remove the #) and add --url to walk the app.
Example: you change one form, and a page you never opened now shows a server error (HTTP 500). The walk reaches the pages your links lead to, so it can find that error before users do.
--url, or a dev server on a common port.The 12 classes:
With no app running, sniff reads your source instead (sniff scan). It flags debug statements,
placeholder text and broken imports, and tells you how to start the real walk.
The MCP server gives your agent these tools:
| Tool | What it does |
|---|---|
sniff | One entry point: mode: "walk", "scan" or "report" |
sniff_scan | Source-only scan, no browser |
sniff_run | Source scan plus browser checks |
sniff_discover | Writes test paths from your source and runs them |
sniff_install | Downloads the engine and Chromium, with progress |
sniff_report | Reads the last saved results |
--caps scan,report limits the server to sniff_scan, sniff_report and the main sniff tool. Then it uses no browser at all.
Each row was tested on 2026-11-17. The first five used the sniff 0.9.0 build, the install rows used 0.9.4.
| Agent | What ran | Tested |
|---|---|---|
| Claude Code | claude -p with sniff as an MCP server called sniff_scan and read 2 findings | 2026-11-17 |
| Codex CLI | codex exec with sniff as an MCP server called sniff_scan and read 2 findings | 2026-11-17 |
| Gemini CLI | gemini extensions validate . accepted gemini-extension.json | 2026-11-17 |
| Any MCP client | node sniff-tests/mcp-smoke.mjs listed 6 tools and a walk returned 20 findings | 2026-11-17 |
| Skills CLI | npx skills@1.9.0 add Aboudjem/sniff installed the 3 skills into a temp folder | 2026-11-17 |
| Claude Code | claude mcp add -s user wrote sniff-qa into a throwaway home's .claude.json | 2026-11-17 |
| Codex CLI | codex mcp add wrote sniff-qa into a throwaway home's config.toml | 2026-11-17 |
| Gemini CLI | gemini extensions install installed and enabled sniff in a throwaway home | 2026-11-17 |
| Cursor | skills add -g -a cursor put the 3 skills in a throwaway ~/.agents/skills | 2026-11-17 |
| GitHub Copilot | skills add -g -a github-copilot put the 3 skills in ~/.agents/skills | 2026-11-17 |
| OpenCode | skills add -g -a opencode put the 3 skills in ~/.agents/skills | 2026-11-17 |
| Claude Code | skills add -g -a claude-code put the 3 skills in ~/.claude/skills | 2026-11-17 |
| Codex CLI | skills add -g -a codex put the 3 skills in ~/.agents/skills | 2026-11-17 |
| Gemini CLI | skills add -g -a gemini-cli put the 3 skills in ~/.agents/skills | 2026-11-17 |
| Skills CLI | skills add -g with no terminal installed into every agent folder | 2026-11-17 |
Documented, not tested today: Windsurf and Zed. Cursor, Copilot and OpenCode were install tested only. Their snippets are in docs/editors.md.
These commands re-print the numbers on this page. The first prints the score on the test app.
npm run build && npm run proof
It starts the planted-bug test app, walks it, and scores the report. Output on 2026-11-17:
Recall: 21/21 (100%)
Findings: 27 deduped (27 raw)
Matched: 27 Candidate FP: 0 Precision proxy: 100%
Hard FP on clean page (/clean): 0
The sample above shows 26 findings and says 1 low-confidence finding is hidden. The proof run uses --all, so it counts all 27.
The test app holds 21 bugs added on purpose (planted bugs) across all 12 classes. It also has one clean page with no bug. The scorer is sniff-tests/score-fixture.mjs. The bugs are listed in sniff-tests/planted-bugs/MANIFEST.json.
To get the same report as the sample, start the test app. Then run sniff in a second terminal:
node sniff-tests/planted-bugs/server.mjs 4517
npx sniff-qa --url http://localhost:4517
npm test
Prints Test Files 54 passed (54) and Tests 603 passed (603) on 2026-11-17.
The suite includes a gate on the test app. It fails the build if sniff finds fewer than 85% of the planted bugs.
It also fails if sniff reports anything on the clean page.
node sniff-tests/mcp-smoke.mjs
Prints the 6 MCP tools and the 20 findings quoted in the table above.
Do I need an API key or an account? No. The walk and the source scan run on your machine. Nothing is sent anywhere unless you set up the optional AI exploration yourself.
Will a walk change data in my app? It can. A walk clicks buttons and submits real forms, so point it at a dev or staging app.
Why did it exit 1 when nothing crashed?
Exit 1 means it found bugs at a failing severity, so CI fails the build. Pass --fail-on none to always exit 0.
Will it edit my code?
No. Walking and scanning never touch your source. Only sniff fix edits files, and only when you run it.
Can it walk a logged-in app?
Yes. Pass --storage-state auth.json. Cookie and token values from that file are hidden in the written reports.
More answers are in docs/faq.md.
needsSetup and the agent calls sniff_install first.sniff.configBuilt by Adam Boudjemaa · Apache 2.0 · built on Playwright and axe-core