Give your AI a
live internet connection
AI models are intelligent, but their training data is stale. When their knowledge runs out, they hallucinate confidently. Grapnel solves that knowledge gap by giving your agents real-time internet access. Let them browse current docs, verify facts, and debug errors just like a human would. Self-hosted, privacy-first, and built for builders.
┌─ grapnel research ── v1.0 ──────────────────────┐ │ Q: What's the best way to handle async file I/O in Python 3.14? │ → SearXNG: searching 50+ engines... │ → DSPy RLM: reading 3 pages progressively │ → Confidence: 0.91 ≥ 0.85 → answer ready │ │ Found: Python 3.14's aiofiles v24.1.0 + anyio v4.6.0 │ Sources: docs.python.org . pypi.org . realpython.com └─ 3 pages read . 2.4s . 0 hallucinations
Smart reading. Zero waste.
Every optimization is designed to give the LM exactly what it needs - nothing more.
Progressive Reading
Pages are read in sections. The LM stops as soon as confidence ≥ 0.85 - no need to finish the page.
Smart Page Fetching
Cache stores full page text. Section-based pagination lets the LM read long pages chunk by chunk without re-fetching.
Cross-Page Confidence
Reads one page at a time and stops when confidence ≥ 0.85. No need to burn tokens on pages you already have the answer from.
Registry-First Versions
7 registries checked in parallel. 200ms response. Zero LLM cost. Falls back to web search only when needed.
Privacy-First Search
SearXNG meta-search across 50+ engines. No Google API. No tracking. Fully self-hosted.
Section Pagination
Long pages are split into sections. Each section is one cache hit. No re-fetching, no content loss.
10 tools. One purpose.
Each tool is designed for a specific task - search, read, verify, debug, research.
A pipeline, not a prompt.
Every query goes through a multi-stage pipeline designed for efficiency.
Deploy in under 5 minutes
Docker Compose handles everything - SearXNG, Valkey cache, and the MCP server.
Clone & Configure
git clone the repo and copy .env.example to .env.
Set API Key
Choose your LLM provider - OpenAI, Anthropic, Gemini, or local Ollama.
Launch
docker compose up -d - starts all three services.
Connect
Point your AI assistant (Cursor, Claude Desktop, etc.) to localhost:8881.
Built on peer-reviewed papers
Grapnel's design is directly motivated by two recent publications.
Recursive Language Models (arXiv:2512.24601)
Shows LLMs can process inputs orders of magnitude beyond context windows by recursively examining information through external tools. The research tool implements this directly.
Is Grep All You Need? (arXiv:2605.15184)
Proves simple keyword-based retrieval outperforms vector search in agentic contexts. Validates SearXNG keyword search over vector databases.
Your AI has a knowledge cutoff.
Grapnel is the bridge.
LLMs don't know when their knowledge is outdated. They answer with confidence - even when they're wrong. Give your agents a live connection to the web. Self-hosted. Privacy-first. Open source.