Plumbref adds a local evidence-checking layer to coding agents
Facundo Taboada built Plumbref, a local-first MCP server that lets coding agents verify their own repo answers by breaking claims into checkable units and classifying each as supported, contradicted, or uncertain before responding.
Score breakdown
Plumbref offloads the verification burden from the user to the agent itself, replacing the manual "are you sure?" follow-up loop with a structured, locally-run claim-checking step built into the MCP workflow.
- 01Plumbref is a local-first MCP server that helps coding agents verify repo answers before delivering them.
- 02Agents use it to break answers into claims, search the repo, read bounded snippets, and check for contradictions.
- 03Each claim is classified as one of six statuses: `supported`, `too_broad`, `uncertain`, `contradicted`, `not_found`, or `not_verifiable`.
Facundo Taboada describes a familiar loop when working with coding agents: the agent produces a confident answer about a repository, but the answer lacks a visible verification trail, prompting repeated follow-up questions like "Did you check the other path?" or "Is this actually only used there?" Plumbref is his response — a local-first MCP server designed to sit between the agent and its final answer, giving the agent a structured process for self-verification before responding.
The core workflow has the agent break its answer into individual claims, search the repository using ripgrep-based search, read bounded source snippets, and check for contradictions.
The core workflow has the agent break its answer into individual claims, search the repository using ripgrep-based search, read bounded source snippets, and check for contradictions. Each claim is then classified into one of six categories: `supported`, `too_broad`, `uncertain`, `contradicted`, `not_found`, or `not_verifiable`. The tool produces Markdown and JSON reports and includes built-in templates covering common verification scenarios such as flow explanations, field migrations, change impact analysis, downstream consumers, and external integrations.
Plumbref is designed to run entirely locally — no hosted service, no repository upload, no database, no vector store, and no external model API key. Taboada notes the project is still early-stage and has been testing it on its own repository, and is seeking feedback from developers using coding agents on real codebases. The source code is available at github.com/ftaboadac/plumbref.
Key facts
- 01Plumbref is a local-first MCP server that helps coding agents verify repo answers before delivering them.
- 02Agents use it to break answers into claims, search the repo, read bounded snippets, and check for contradictions.
- 03Each claim is classified as one of six statuses: `supported`, `too_broad`, `uncertain`, `contradicted`, `not_found`, or `not_verifiable`.
- 04Runs entirely on the user's machine — no hosted service, no repo upload, no database, no vector store, no external model API key.
- 05Ships with a CLI, Markdown/JSON report output, and ripgrep-based repository search.
- 06Includes built-in templates for flow explanations, field migrations, change impact, downstream consumers, and external integrations.
- 07The project is early-stage; Taboada has been dogfooding it on its own repository.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 13, 2026 · 08:58 UTC. How this works →