MCP server

Dungbeetle ships an MCP server — dungbeetle-mcp — so coding agents can triage snapshot runs and approve baselines directly, without leaving their tool loop.

It is self-hostable and runs against your own Dungbeetle server, so nothing is sent to a third-party cloud. The design follows one principle: agents get low-token structured diff summaries (what changed, where, how severe), never base64 screenshots — images stay behind a review URL a human opens.

Deterministic diffs, human-owned approval. The agent reads and triages; mutating a baseline is a destructiveHint tool, so your MCP client asks for human confirmation before it runs.

What it talks to

dungbeetle-mcp is a thin client over the server's /api/v1 surface and authenticates with a repository's client credentials — the same client_id / client_secret pair dungbeetle push uses. Everything it can see or change is scoped to that one repository.

Capture stays in the CLI (dungbeetle ci → dungbeetle push); the MCP server is the review and approval surface over the runs you've pushed.

Next: set it up, then the tools and agent loop.