Technical SEO that fixes itself.
An agent for your AI coding client. Point it at your site: it finds what's broken, fixes it in your repo, and proves it's live.
# add Searchlight to your AI coding client claude mcp add searchlight -- npx -y @ajmalaksar/searchlight serve # then just ask "why isn't my site showing up on Google? fix it."
From “why isn't Google indexing me?” to fixed and deployed.
Most tools stop at the diagnosis. Searchlight runs the whole loop and proves the last step.
Detect
Reads your real Search Console and Analytics data and crawls your key pages.
Explain
Triages every finding in plain language: fix now, worth improving, or safe to ignore.
Fix
Edits your repository, framework-aware — canonical, redirects, sitemaps, metadata, the tag.
Deploy
Commits and ships through your pipeline. The deploy isn't handed back to you.
Verify
Re-audits the live site in a real browser: the redirect resolves, the tag fires.
One real site, the whole run.
A live session on zawaaj.in — a custom Next.js site with a host and canonical conflict and missing analytics. Not a summary; the actual loop.
Homepage audit 90/100. Both https://zawaaj.in and https://www.zawaaj.in return 200 with no canonical, so Google reports Duplicate without user-selected canonical. Meta description 198 chars (truncates). No GA4 tag on the page.
Fix now: the apex and www both resolve, fighting for the canonical. Worth improving: trim the meta description. Fine: sitemap reachable, mobile usability clean.
One canonical host, enforced in code and agreed by the redirect.
git push → Vercel build → live. Most of this step is waiting on the platform, not work.
curl -I https://www.zawaaj.in → 308 → https://zawaaj.in 200; GA4 G-… fired in-browser.
audit 90 → 98 · canonical set · redirect www→apex · meta 198 → 155 · GA4 live · sitemap submitted
How to read this. The before/after state and the ~50-minute wall-clock are measured; the manual hours are typical estimates for the same scope, not a stopwatch. Searchlight automates the diagnosis and the fix (canonical, host, redirects, sitemap, the GA4 install and verification). It does not design your ecommerce event-tracking plan; that part is still a human's job.
One agent for the whole technical surface.
Search Console tells you what is wrong. Searchlight works out what to change, makes the change, and proves it worked.
Indexing and coverage
Which pages Google has indexed, which it hasn't, and why.
Canonical and host conflicts
The www-vs-apex split that fights your ranking, aligned to one host.
Redirect chains
Loops and chains collapsed to one clean hop.
Sitemaps that resolve
A sitemap that returns 200, with the right URLs, submitted.
GA4 set up, not pasted
The tag installed and firing. No copy-pasting IDs.
Core Web Vitals
LCP, INP and CLS from real-user field data.
On-page audits
Titles, meta, canonical, OG and schema, as concrete edits.
Scores and reportin progress
A 'fix these first' list and a shareable report.
Add it to your client. The agent does the rest.
An MCP server with a companion /searchlight skill. Add both, sign in once, then ask.
Add the server and the skill
One command adds the MCP server; one more installs the /searchlight skill.
claude mcp add searchlight -- npx -y @ajmalaksar/searchlight servesearchlight skill installSign in with Google
One local OAuth sign-in for Search Console and Analytics. The token stays on your machine.
searchlight login --setupAsk
Run the loop in plain language. It reports back with the before and after.
/searchlight audit yoursite.comYour data never leaves your machine.
Searchlight runs as a local server. You sign in with your own Google account. No hosted backend, no warehouse.
- Read-only Search Console and Analytics scopes for diagnosis. Write and provisioning scopes are opt-in, requested only when you ask it to set things up.
- OAuth tokens are stored locally under
~/.searchlightand never transmitted to us. - No hosted backend and no shared database. Your site's data stays yours.
- Open source and MIT licensed. Read every line.
See what is actually broken on your site.
Point it at a site, get a triaged, plain-language picture in minutes, then let it fix the things worth fixing.