MCP for coding agents
Connect Claude Code, Cursor or any MCP client to webtrax mcp, so your agent checks your dev server before it commits.
webtrax mcp runs a local MCP server over stdio. Your coding agent can check the pages it just changed for broken links, redirects, security headers and SEO problems, and see only what its change broke.
Connect it
Claude Code:
claude mcp add webtrax -- npx -y @webtrax/cli mcp
Cursor (.cursor/mcp.json), and most other clients:
{ "mcpServers": { "webtrax": { "command": "npx", "args": ["-y", "@webtrax/cli", "mcp"] } } }
Runs are saved in .webtrax/runs in the directory the server starts in, shared with the CLI.
Tools
| Tool | What it does |
|---|---|
check_urls |
Check up to 200 URLs (for example links the agent just wrote) without crawling: ok, broken, redirect, timeout, dns_error, rate_limited or blocked. |
audit_page |
Check one page and its links with every suite; returns findings with fixes, plus the page’s title, description, headings and word count. |
start_scan |
Crawl a site in the background; returns a run_id. |
get_scan |
Status, progress and summary of a run. |
wait_for_scan |
Wait up to 120 seconds for a run to finish, with progress notifications. |
list_findings |
Findings of a run (default: the latest), filtered by minimum severity or suite, paged. new_only: true shows only findings that were not in the previous run. |
get_finding |
One finding in full: evidence, every page it was seen on, and the fix. |
compare_scans |
New, fixed and persisting findings between two runs. |
explain_finding |
Why a finding or rule matters and how to fix it. Never calls an external service. |
It also offers the latest report as a resource (webtrax://local/runs/latest/report.md), each rule as webtrax://local/rules/{rule_id}, and a pre-commit-site-check prompt.
Safety
- Local targets only by default: localhost,
*.localhost,*.local,*.test,*.internal, loopback and private IPs. A prompt-injected agent cannot turn your laptop into a crawler of other people’s sites. Start the server with--allow-publicto allow public sites; crawls are then capped at 500 pages and 2 requests per second per host. - Page content is untrusted. Results that include text read from a website are marked
untrusted_content: true, because a page can carry instructions aimed at your agent. Treat them as data. - No telemetry. The server talks only to the sites you ask it to check.