claim MCP server string-matches citations against sources and flags fabricated quotes
BondableColt18 · reddit · 2026-10-23
A developer shares claim, an open-source MCP server built for a recurring pain point: a model's citation looks perfect, but the quote simply isn't in the file.
- Verification uses string matching rather than asking an LLM to judge plausibility: checking a quote returns a page image highlighted green for full match, amber for partial, red for none.
- It can scan a full answer against a set of sources: every shared run of eight words and every figure is checked; numbers found in no file are flagged as "computed" — derived, not sourced.
- Supports PDF, XLSX/CSV/TSV, text/markdown, images and web pages; ships as a Claude Code plugin that auto-scans in-session answers, and the MCP server works standalone with any client. Runs fully locally — no uploads, no accounts.
Open-sourced at GitHub (alnaik/claim); the author seeks feedback on edge cases where the 8-word overlap or shared-words rule for numbers breaks down.
More from coding & agent
- Keeping multiple AI agents productive is more taxing than any programming, devs say — generativist · 2026-10-23
- Vibecoding your own LLM inference engine: GGUF in WebGL2 ends in debug hell — CodeMichaelD · 2026-10-23
- Dev builds an agent-human 'software factory' against full-AI delivery — Pavan_Belagatti · 2026-10-23
- Open-source Harness Engineering repo turns coding-agent guidance into mechanical guardrails for Claude Code — blaizedsouza · 2026-10-23
- Harness Engineering: How OpenAI and LangChain Got Agents to Ship a Million Lines of Code — blaizedsouza · 2026-10-23
- Agent Relay, a YC-backed open-source tool, lets your AI agents talk to each other directly — ycombinator · 2026-10-23