██╗     ███████╗ █████╗ ███╗   ██╗     ██████╗████████╗██╗     
  ██║     ██╔════╝██╔══██╗████╗  ██║    ██╔════╝╚══██╔══╝██║     
  ██║     █████╗  ███████║██╔██╗ ██║    ██║        ██║   ██║     
  ██║     ██╔══╝  ██╔══██║██║╚██╗██║    ██║        ██║   ██║     
  ███████╗███████╗██║  ██║██║ ╚████║    ╚██████╗   ██║   ███████╗
  ╚══════╝╚══════╝╚═╝  ╚═╝╚═╝  ╚═══╝     ╚═════╝   ╚═╝   ╚══════╝
          Context-Engineered AI Coding CLI

The Only Coding CLI That Steers How Much Your LLM Thinks — Powered by lean-ctx

40+ Tools · Thinking Steering · Hybrid Semantic Search · Symbol Intelligence · Multi-Provider

Website · Install · How It Works · Features · Providers · Changelog · Discord

---

leanctl is a terminal-native AI coding agent that automatically steers LLM thinking budgets based on task complexity — so simple tasks stay fast and cheap, while complex tasks get the deep reasoning they need.

⚡ How Thinking Steering Works

  User: "fix the typo in line 42"
    → Intent: TaskType::QuickFix
    → Budget: Minimal (1024 tokens)
    → Result: Fast, cheap, accurate

  User: "debug why the distributed cache eviction is wrong"
    → Intent: TaskType::Debug
    → Budget: High (8192 tokens)
    → Result: Deep analysis, thorough investigation

  User: "redesign the auth architecture for microservices"
    → Intent: TaskType::Architecture
    → Budget: Maximum (16384 tokens)
    → Result: Comprehensive design with trade-offs

Task Type	Thinking Budget	Token Savings vs Max
Quick fixes, renames, formatting	Minimal (1K)	94%
Bug fixes, small features	Low (2K)	88%
Debugging, refactoring	Medium (4K)	75%
Architecture, complex analysis	High-Max (8-16K)	Optimal

Lab-proven: Thinking steering reduced output tokens by 54-95% and thinking tokens by 48-100%.

📦 Install

# macOS / Linux — one-liner
curl -fsSL https://leanctl.com/install.sh | sh

# npm (all platforms)
npm install -g leanctl-bin

# Arch Linux (AUR)
yay -S leanctl-bin

# Upgrade existing installation
leanctl upgrade

🚀 Quick Start

# 1. Configure your AI provider (interactive)
leanctl init

# 2. Start coding
cd your-project
leanctl

🧠 Features

40+ LLM-Accessible Tools

Category	Tools	What They Do
File Intelligence	read_file, file_outline, symbol_lookup	Read files, get outlines, look up symbols by name
Code Navigation	find_callers, find_callees, search_code	Navigate call graphs, find references
Semantic Search	semantic_search (hybrid)	BM25 + dense vector search with language filters
Architecture	extract_routes, graph_diagram, related_files	HTTP endpoints, Mermaid diagrams, dependency analysis
Editing	write_file, edit_file, multi_edit	Precise code modifications
Shell	run_command	Compressed output with 90+ patterns
Memory	project_knowledge, compress_memory	Cross-session knowledge, markdown compression
Multi-Agent	agent_coordination, task	Agent bus, sub-agent spawning

Context Engine (lean-ctx v3.0.2)

• 8 read modes — full, map, signatures, diff, aggressive, entropy, lines:N-M, reference
• Session caching — re-reads cost ~13 tokens instead of thousands
• 90+ compression patterns — git, npm, cargo, docker, test runners
• Adaptive compression — Thompson Sampling bandits learn optimal settings per language

Multi-Provider Support

Provider	Models	Thinking Steering
Anthropic	Claude 4, Claude 3.5 Sonnet	✅ budget_tokens
OpenAI	GPT-4.1, o3, o4-mini	✅ reasoning_effort
Google	Gemini 2.5 Pro/Flash	✅ thinking_budget
Ollama	Any local model	—

📥 Downloads

Pre-built binaries for every release:

Platform	Binary	Checksum
macOS Apple Silicon	leanctl-aarch64-apple-darwin.tar.gz	✅ SHA256
macOS Intel	leanctl-x86_64-apple-darwin.tar.gz	✅ SHA256
Linux x86_64	leanctl-x86_64-unknown-linux-musl.tar.gz	✅ SHA256

Linux binaries are statically linked (musl) — they work on every Linux distribution.

🔗 Links

• Website: leanctl.com
• Core Engine: lean-ctx — MIT-licensed, fully open source
• Discord: Join the community
• X/Twitter: @leanctx

📜 License

LeanCTL is proprietary software. The binary is free to use for personal and commercial projects.

The underlying engine lean-ctx is MIT-licensed and fully open source.