Claude Code

Give Claude Code the Dungbeetle MCP tools so it can triage snapshot runs and — with your confirmation — approve baselines. You stay the account owner; Claude holds only a scoped, revocable agent token.

1. Prerequisites

• The Dungbeetle CLI: npm install -g dungbeetle.
• Node.js with npx (the server is on npm as dungbeetle-mcp — see Setup).
• A Dungbeetle cloud account with a repository (getting access).

2. Mint a token

dungbeetle login --label "Claude Code"

Approve the printed code at /ui/connect in your browser, picking the repository Claude may work with. Or approve the code from any signed-in browser on another machine — see Agent authentication.

3. Add the MCP server

claude mcp add dungbeetle \
  --env DUNGBEETLE_AGENT_TOKEN=dbat_… \
  -- npx -y dungbeetle-mcp

Or commit a project-scoped .mcp.json at the repo root so every teammate's Claude Code picks it up (each teammate supplies their own token — don't commit one):

{
  "mcpServers": {
    "dungbeetle": {
      "command": "npx",
      "args": ["-y", "dungbeetle-mcp"],
      "env": {
        "DUNGBEETLE_AGENT_TOKEN": "${DUNGBEETLE_AGENT_TOKEN}"
      }
    }
  }
}

Verify with claude mcp list.

Self-hosting

Both commands default to the hosted cloud. On a self-hosted server, pass --server <url> to dungbeetle login and add "DUNGBEETLE_SERVER_URL": "https://your-server" to the env block above.

4. First loop

Ask Claude Code:

List the recent Dungbeetle runs and show me the semantic diff of the latest failed one.

Claude calls dungbeetle_list_runs, then dungbeetle_get_run with detail: "full" — a compact node-level diff, never screenshots. Fix the regression (or decide the change is intended), then:

Approve that run and promote the new baselines.

dungbeetle_approve_run is a destructive tool, so Claude Code asks for your confirmation before it runs — approval stays human-owned.

Scopes

Use the Triage preset (runs:read baselines:read analytics:read reviews:write) — the default when dungbeetle login requests no explicit scopes. Claude can read diffs and record your review decisions but can't push runs or rewrite baselines outside the promote-on-approve path. See scopes.