llm-tools
The missing standard library for Agentic Workflows. Native Go. Single Binary. 100x Faster than Python.
⚡ Why this exists
LLM Agents need to be fast. Waiting 400ms for a Python script or 100ms for Node.js to spin up just to read a file kills the flow of an autonomous loop.
llm-tools is a suite of high-performance, statically compiled tools designed to be the "hands" of your AI agent. It includes a native MCP Server for instant integration with Claude Desktop and Gemini.
The "Rewrite it in Go" Effect
I benchmarked this against equivalent Python and Node.js implementations on a real-world codebase. The difference is massive.
vs Python (llm-support)
Benchmarks run on M4 Pro 64gb macOS Darwin (arm64), 2025-12-26.
vs Node.js (llm-filesystem)
I ported the popular fast-filesystem-mcp from TypeScript to Go to create llm-filesystem.
Benchmarks run on M4 Pro 64gb macOS Darwin (arm64), 2025-12-31.
🚫 Zero Dependency Hell
Deploying agent tools in Python or Node is painful. You have to manage virtual environments, node_modules, pip install dependencies, and worry about version conflicts.
llm-tools is a single static binary. It works instantly on any machine—no setup required.
🤖 Standardized LLM Orchestration
llm-tools isn't just for reading files; it's a reliability layer for your agent's cognitive functions.
Two ways to call LLMs:
prompt - Universal CLI Adapter
Wraps external LLM CLIs (gemini, claude, ollama, openai, octo) with retries, caching, and validation:
llm-support prompt \
--prompt "Analyze this error log" \
--llm gemini \
--retries 3 \
--cache \
--min-length 50
complete - Direct API Access
Calls OpenAI-compatible APIs directly without external binaries:
# Uses OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL
llm-support complete \
--prompt "Explain this code" \
--system "You are a senior engineer" \
--temperature 0.3
⚡ Advanced Workflows
Parallel Batch Processing (foreach)
Run prompts across thousands of files in parallel without writing a loop script. Perfect for migrations, code reviews, or documentation generation.
# Review all Go files in parallel (4 concurrent workers)
llm-support foreach \
--glob "src/**/*.go" \
--template templates/code-review.md \
--llm claude \
--parallel 4 \
--output-dir ./reviews
🚀 Quick Start
Pre-built Binaries
Recommended: Download the latest binary for your OS. No dependencies required.
Installation (Go)
go install github.com/samestrin/llm-tools/cmd/llm-support@latest
go install github.com/samestrin/llm-tools/cmd/llm-filesystem@latest
go install github.com/samestrin/llm-tools/cmd/llm-semantic@latest
💡 Common Recipes
See what's possible with a single line of code:
# Find all TODOs and FIXMEs (Fast grep)
llm-support grep "TODO|FIXME" . -i -n
# Show project structure (3 levels deep)
llm-support tree --path . --depth 3
# Search for multiple definitions in parallel (Token optimized)
llm-support multigrep --path src/ --keywords "handleSubmit,validateForm" -d
# Extract data from JSON without jq
llm-support json query response.json ".users[0]"
# Calculate values safely
llm-support math "round(42/100 * 75, 2)"
# Generate config from template
llm-support template config.tpl --var domain=example.com --var port=8080
# Hash all Go files (Integrity check)
llm-support hash internal/**/*.go -a sha256
# Count completed tasks in a sprint plan
llm-support count --mode checkboxes --path sprint/plan.md -r
# Detect project stack
llm-support detect --path .
# Extract only relevant context (AI-filtered) - works with files, dirs, and URLs
llm-support extract-relevant --path docs/ --context "Authentication Config"
llm-support extract-relevant --path https://docs.example.com --context "API keys"
# Fetch a URL as Markdown, with proxy + FlareSolverr fallback (JS via --via flaresolverr)
llm-support fetch https://example.com --format markdown
# Extract and rank links from any webpage (heuristic scoring)
llm-support extract-links --url https://example.com/docs --json
# Extract links with LLM-based relevance scoring
llm-support extract-links --url https://example.com/docs --context "authentication" --json
# Summarize directory content for context window (Token optimized)
llm-support summarize-dir src/ --format outline --max-tokens 2000
# Batch process files with a template (LLM-driven)
llm-support foreach --files "*.ts" --template refactor.md --parallel 4
📚 Documentation
Detailed references for all 40+ commands:
- llm-support Commands - File ops, search, analysis, and data processing.
- llm-filesystem Commands - High-performance filesystem operations (drop-in replacement for fast-filesystem-mcp).
- llm-semantic Commands - Semantic code search with local embeddings.
- MCP Setup Guide - Integration with Claude Desktop & Gemini.
- Quick Reference - Cheat sheet.
🧠 How It Works
The Loop:
- Agent receives a task.
llm-supportprovides fast codebase context (files, structure, search results).llm-semantic memoryrecalls past decisions ("Use Jest, not Mocha") to prevent regression.- Agent generates code with full context.
sequenceDiagram
participant U as 👤 User
participant A as 🤖 Agent
participant S as ⚡ Support
participant M as 🧠 Memory
participant C as 📂 Codebase
U->>A: /execute-sprint
rect rgb(30, 30, 30)
note right of A: Fast Context
A->>S: multiexists, count, report
S-->>A: ✓ Context Loaded (22ms)
end
rect rgb(50, 20, 20)
note right of A: Long-Term Memory
A->>M: memory search
M-->>A: ⚠ RECALL: "Use Jest"
end
A->>C: TDD Implementation (using Jest)
