io.github.4DA-Systems/4da-mcp-server

All signal. No feed.

4DA reads the internet for developers — privately, locally — and gets sharper every day.

It scans your codebase — Cargo.toml, package.json, go.mod, Git history — and scores every article, advisory, and release from 20+ sources against what you actually build. An item needs 2+ independent signals to survive. Everything else is rejected.

Tested across 9 developer personas: 92% of content is filtered as noise, 98% of actual noise is correctly rejected. Your real rejection rate — computed from your own data — is shown in the Evidence tab.

It learns from how you engage with what it shows you. Save something — topics boost, source reputation rises, your taste embedding sharpens. Dismiss something — anti-patterns form, future noise drops. Yesterday's noise becomes tomorrow's signal.

The fastest way to try it

Already using Claude Code, Cursor, or Windsurf? One command:

npx @4da/mcp-server

This scans your project, detects your stack, and gives your AI assistant live vulnerability scanning, dependency health, upgrade planning, and ecosystem intelligence. No API keys. No accounts. Works standalone — no desktop app required. Full MCP documentation.

How It Works

Scoring

5 independent signal axes. An item must pass 2 or more to surface. Single-axis matches are hard-capped at 28% — no matter how strong one signal is, it cannot pass alone.

What passes the gate goes through 12 quality multipliers: content depth, novelty detection, competing tech penalties, title-body coherence, and intent scoring from recent work. Every constant is calibrated across 9 simulated developer personas with 215 labeled test items.

LLM Verification

After keyword scoring, an LLM layer verifies the top items against your full developer context — stack, dependencies, recent commits, anti-technologies, and engagement history. Strict 1-5 rubric:

• 5 = MUST-READ: Security alert for YOUR dependency, breaking change YOU must act on
• 3 = WORTH KNOWING: Useful tool that fits YOUR exact stack
• 1 = NOISE: Mentions your tech but isn't actionable

This is where the gold surfaces — articles the keyword pipeline misses because there's no keyword overlap, but the LLM understands the conceptual relevance to your specific project.

You own the compute. Use Ollama for free local inference (fully private), or bring your own Anthropic/OpenAI key. 4DA never pays for your compute, never stores your keys remotely, never makes API calls you didn't configure.

Anti-Gaming

Content creators who learn the scoring algorithm still can't game it:

• Title-body coherence: titles must deliver on what they promise. Claim "React + Rust + Tauri" but only discuss React? Penalty.
• **Keyword concentratio

... View full README on GitHub