jev-research-mcp

Stop burning tokens with unrelated search results,filter em!

Developer Tools

Artificial Intelligence

GitHub

Vibe coding

I build with coding agents every day, and I kept watching the same wasteful pattern: Agent does web research -> fetches 20 full pages -> dumps navigation, cookie banners, ads, and duplicate docs straight into the most expensive model in the loop. It was slow, it burned context, and half of what the model was reading was junk it would never use. JEV Evidence MCP is my fix for that. It's a self-hosted, open-source MCP server with one tool: research(query). Instead of returning raw pages, it filters twice with TypeSafe JEV: 1. Before fetching: Is this search result relevant + useful? If not, we never fetch it at all - no network work for junk. 2. After parsing: Is this block relevant + actionable for a coding agent? If not, we cut it. What you get back is compact evidence blocks with provenance - verbatim code, headings restored, URL attached. No rewriting, no summarizing, no hallucinated citations. The downstream model still does the reasoning, it just gets clean material to reason over. Privacy was a core motivator too. This is local-first, BYOK. Your keys, queries, and docs stay on your machine except to Brave (for search), TypeSafe (for judgments), and the source URLs you chose. I receive nothing. No telemetry, no hosted service. How did your approach or process evolve while working on this launch? Three big shifts: 1. From generation to selection. My first instinct was "summarize the web." I killed that fast. Summarizers destroy code formatting and invent facts. The principle became: JEV decides, code disposes. JEV answers semantic yes/no questions, deterministic code does everything else - HTTP, parsing, batching, budgets, dedup. We select evidence, we never generate it. 2. From aggressive filtering to conservative retention. Early versions cut too much. I learned false negatives hurt way more than false positives - dropping the only right doc breaks the agent. So now if JEV is uncertain, we keep it. The policy is explicit: keep = relevant AND useful, with an uncertain band that is retained by default. All thresholds are configurable. 3. From claims to proof. I refused to ship with "90% fewer tokens!" marketing. So I built a baseline mode (no JEV) and a live benchmark that compares baseline vs. JEV on identical queries - fetches avoided, blocks seen/kept, tokens, latency. No fixture numbers. Run it on your workload and quote your own numbers. I also added a localhost dashboard because the funnel is the product - you can literally watch search -> selected -> fetched -> blocks seen/kept -> evidence. If you can't see what was cut and why, you won't trust it. That's what I'm launching: filter twice, fetch less.