I built Plumbref because I kept running into the same loop with coding agents
They would give a confident answer about a repo, and then I’d have to ask:
- “Are you sure?”
- “Did you check the other path?”
- “What about downstream consumers?”
- “Is this actually only used there?”
The issue wasn’t that the agent couldn’t read code. The issue was that the answer didn’t always come with a clear verification trail
Plumbref is a local-first MCP server that helps a coding agent check repo answers before giving them back
It lets the agent:
- break an answer into claims
- search the repository
- read bounded source snippets
- check for contradictions
- classify claims as supported, too_broad, uncertain, contradicted, not_found, or not_verifiable
Plumbref runs locally:
- no hosted service
- no repo upload
- no database
- no vector store
- no model API key
Right now it includes an MCP server, a CLI, Markdown/JSON reports, ripgrep-based repo search, bounded snippets, broad-claim detection, and built-in templates for flow explanations, field migrations, change impact, downstream consumers, and external integrations
It’s still early, but I’ve been dogfooding it on its own repo and I’d like feedback from people using coding agents on real codebases
Top comments (2)
Love the evidence-checking angle — verification before output is exactly what agents need. There's a parallel problem on the input side: agents don't verify whether they should be thinking or acting before making tool calls. I put together Brainstorm-Mode (mehmetcanfarsak on GitHub) that adds a hook-based gate — PreToolUse intercepts execution during brainstorming phases so the agent actually produces ideas instead of premature code. Same philosophy as your checker, just on the ideation side.
Thanks! Nice parallel, I like the idea of gating when agents should act and not just checking what they output. I’ll check out Brainstorm-Mode!