Skip to content

AI agents (MCP)

Overwatch runs a remote Model Context Protocol server. An AI agent (Claude, Cursor, Codex, or any MCP client) connects to it, reads what’s wrong with a site, and fixes it where the site lives: in its code or its CMS.

Overwatch reports and suggests. The agent does the fixing. Every finding carries why it matters and a suggested fix, plus enough detail (the link target, a CSS selector, the offending text) to find it.

In the dashboard, Organization → AI agents has step-by-step instructions for Claude Code, Claude, Cursor, VS Code, Codex and ChatGPT. The server URL is:

https://api.overwatch.weareheavy.dev/mcp

Add it to the agent as a remote (HTTP) MCP server. For Claude Code:

Terminal window
claude mcp add --transport http overwatch https://api.overwatch.weareheavy.dev/mcp

The first call opens Overwatch in the browser:

  1. Sign in, if you aren’t already.
  2. Choose the organization the agent may read. One connection reads one organization.
  3. Review what the agent may do, and Allow.

The agent sees exactly the sites you see in that organization. If your access to a site, or to the organization, is removed, the agent loses it on its next call.

Organization → AI agents lists the connected agents, with who connected each one, what it may do, when it was last used and how many calls it made. Owners and admins see everyone’s agents; members see their own.

Disconnect cuts an agent off on its next call. To use it again, connect it again from the agent. Owners and admins can disconnect anyone’s agent; members can disconnect their own.

Tool What it answers
list_sites The sites you can see, which checks are on, and when each was last crawled.
get_site_overview What to fix first: totals by severity, the highest-priority pages (weighted by search traffic when Search Console is connected), what changed, and search signals.
get_page_findings Everything known about one page, across broken links, SEO, accessibility and Google.
list_broken_links Broken links from the latest crawl: the page, the target, the status, the link text and a CSS selector.
list_seo_issues SEO issues with their lifecycle (open, resolved, new).
list_accessibility_issues Accessibility findings with impact, WCAG criterion and example elements.
get_seo_summary The site-wide SEO picture: score, crawl health, site structure and sitemap, structured data, Google’s index view, and each rule with its fix.
get_ai_readiness AI readiness: which AI crawlers robots.txt blocks, llms.txt, and AI findings per rule and per page.
get_performance Core Web Vitals (real-user data from Chrome, or a Lighthouse run when there’s none) with the pass/fail verdict, and Lighthouse’s opportunities.
get_search_performance Search Console over 28 days: clicks, impressions, top pages and queries, index coverage, and insights.
get_history Trends across crawls: broken links, SEO, AI readiness and accessibility scores.
get_crawl_status Whether a crawl is running now, the latest finished crawl, and when the next is due.
list_uptime_monitors Uptime monitors and their current status.
get_uptime_report One monitor’s uptime per day and its recent incidents.
get_launch_readiness For a new site replacing a live one: the launch verdict and checklist, and the live URLs that would 404 or end in a broken redirect, with a suggested redirect target.
recheck_page After a fix: re-read the live page now and re-check its links.

Everything except recheck_page is read-only. The agent can’t change settings, sites, members or keys.

The agent asks for these OAuth scopes:

  • overwatch:read: read findings. Every tool needs it.
  • overwatch:recheck: re-check pages. If the agent wasn’t granted it, recheck_page asks you for it the first time it’s used.
  • offline_access: stay connected without asking again.
  • Requests: 120 per minute per organization. Above that, calls fail until the minute is over.
  • Re-checks: 200 per organization per UTC day. Fixed pages are re-checked on the next crawl anyway, so re-check the pages you changed, not the whole site.

Every tool call is recorded per organization, user, agent and day. The AI agents page shows it per agent.

  • Protocol: MCP 2026-07-28 over Streamable HTTP, stateless. 2025-11-25 clients are served on the same endpoint.
  • Authorization: OAuth 2.1 with PKCE. Discover it from the protected resource metadata at /.well-known/oauth-protected-resource/mcp, which points at the authorization server https://api.overwatch.weareheavy.dev/api/auth.
  • Client registration: identify the client with a Client ID Metadata Document: an HTTPS client_id URL that serves client_name and redirect_uris. Dynamic Client Registration is off.
  • Resource indicator: send resource=https://api.overwatch.weareheavy.dev/mcp on the authorization and token requests. Tokens are bound to it.
  • DPoP: supported, not required.