froggeric/llm

Give your local AI coding agent vision and web search — fully on your own machine. No cloud, no API keys, no telemetry.

A monorepo of on-device, privacy-first tools for working with LLMs and coding agents (Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf, and friends): MCP servers that add capabilities text-only models lack, drop-in Claude Code skills, a curated prompt library, and the open-weights VLM benchmark that backs the model choices.

---

🔭 localvision — vision for text-only coding LLMs

The headline tool. localvision is a Go MCP server that wraps a local llama.cpp subprocess so text-only coding agents — Claude Code, Codex CLI, Gemini CLI free tier, non-vision Claude variants — can finally see images.

• Private by design. Images are processed on-device; llama-server binds to 127.0.0.1 only. No telemetry, ever. The only outbound traffic is the one-time model download from huggingface.co/froggeric/.
• 11 task-tuned tools — OCR, code/UI/diagram/chart extraction, PDF ingestion, image-to-prompt, image diff — each routed to its benchmark-best model.
• Cross-platform: macOS (Apple Silicon/Intel), Linux, Windows (x86_64 + arm64).
• Use it two ways — as a one-shot shell command or an MCP server.

Install:

# macOS (Homebrew)
brew tap froggeric/homebrew-tap && brew trust froggeric/tap && brew install localvision

# any OS, no Homebrew:
curl -fsSL https://github.com/froggeric/llm/releases/latest/download/install.sh | bash

# or with Go:
go install github.com/froggeric/llm/mcp/localvision/cmd/localvision@latest

Use it from Claude Code (or any MCP client):

{ "mcpServers": { "localvision": { "command": "localvision", "args": ["run"] } } }

…then ask: "Use the read_image tool to describe ./screenshot.png." Or one-shot from the shell: localvision img.png --type ocr.

➡️ Full docs, latency expectations, and the model catalog: mcp/localvision/README.md

---

🌐 SearXNG MCP — free, unlimited web search for your AI tools

A local SearXNG metasearch + MCP server that gives your coding agent web search and URL reading — free, unlimited, no API key, no Docker. One setup script auto-detects your tools (Claude Code, Gemini CLI, Cursor, Windsurf, Claude Desktop, iFlow, Qwen Code, OpenCode) and wires in the MCP server; a macOS launchd job keeps it running.

➡️ mcp/searxng/readme.md

---

🧠 Claude Code skills

Drop-in skills for Claude Code (copy a folder into your skills directory):

• prompt-enhancer — turns a vague request ("fix the bug", "make it better") into a precise, executable prompt before acting.
• intelligent-versioning — release prep on autopilot: semantic-version bumps, changelog, and documentation sync.
• agentic-translation — an agentic workflow for translating documents of 5K+ words with consistency across passes.

➡️ claude/skills/

---

📝 Prompt library

Curated, shell-centric prompt templates for coding tasks — a prompt-engineer meta-prompt, an introspection step, a competitive code-review prompt, two "Elements of Style" editions, and a web-project starter. Copy-paste ready and easy to combine with shell scripts and CLI clients.

➡️ prompts/coding/

---

📊 VLM benchmark (for researchers)

The open-weights Vision-Language Model benchmark used to choose localvision's catalog: 30 images × 15 model variants × 3 runs at Q4_K_M, plus a Q8_0 comparison. If you're picking a local vision model for your own hardware, the ranking, per-model analysis, and hardware-tier recommendations are a great starting point.

➡️ benchmark/vlm/BENCHMARK-REPORT-v5.md · SUMMARY.md

---

License & privacy

• Source-available under PolyForm Noncommercial 1.0.0 — read, modify, and redistribute for non-commercial use; commercial use requires a separate license. The Claude skills and prompt templates are provided as-is for you to adapt.
• Everything runs on your machine. No telemetry; the only outbound traffic is the one-time model/binary download over HTTPS to huggingface.co/froggeric/. llama-server binds to 127.0.0.1 only.

---

Repository layout

froggeric/llm
├── mcp/
│   ├── localvision/   # vision MCP server (the flagship) — Go, 6 platforms
│   └── searxng/       # local web-search MCP server
├── claude/skills/     # Claude Code skills: prompt-enhancer, versioning, translation
├── prompts/coding/    # curated prompt templates
└── benchmark/vlm/     # open-weights VLM benchmark + reports