comber-mcp

An MCP server for Comber — the every-click QA agent. Point your editor's agent at a preview URL and it walks every inch of the app, then reports the broken clicks, dead ends, console errors, and broken links right inline.

One server, every MCP client: Cursor, Claude Desktop, Claude Code, Windsurf, Zed.

Tools

| Tool | What it does | | --- | --- | | comber_check | Crawl a web app / preview URL and report findings. Read-only by default. | | comber_api_check | Exercise an OpenAPI spec, GraphQL endpoint, or recorded HAR. | | comber_report | Pull the full findings list from a prior run. |

Checks are read-only by default — Comber does not submit forms or trigger mutations unless you explicitly pass allowWrites. The optional LLM finding-judge only runs when you have your own ANTHROPIC_API_KEY in the environment; nothing is billed to Comber.

Install

The server runs straight from npm — no clone, no build.

Cursor

Settings → MCP → Add new server, or drop this in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "comber": {
      "command": "npx",
      "args": ["-y", "comber-mcp"],
      "env": { "ANTHROPIC_API_KEY": "" }
    }
  }
}

Leave ANTHROPIC_API_KEY empty for heuristic-only checks, or set it to enable the LLM judge.

Claude Code

claude mcp add comber -- npx -y comber-mcp

Claude Desktop

In claude_desktop_config.json:

{
  "mcpServers": {
    "comber": { "command": "npx", "args": ["-y", "comber-mcp"] }
  }
}

Windsurf

In ~/.codeium/windsurf/mcp_config.json — same mcpServers shape as above.

Zed

In settings.json:

{
  "context_servers": {
    "comber": { "command": { "path": "npx", "args": ["-y", "comber-mcp"] } }
  }
}

Usage

Once installed, just ask your agent:

Run comber on https://my-branch.preview.app and fix anything that fails.

The agent calls comber_check, gets a coverage + findings summary back, and can call comber_report for the full detail before proposing fixes.

Gate a merge:

Run comber with strict on the preview and block if there are any fail findings.

Notes

• The first web check may prompt Playwright to download a browser; run npx playwright install chromium once if you hit a launch error.
• Runaway-cost backstops (maxStates, maxActions) are exposed as tool inputs.
• Full engine docs: https://codecomber.io.