Claude Code Local (`ccl`) Documentation

Run the official Claude Code CLI against your own local models — anything in Ollama, LM Studio, or a GGUF from Hugging Face — at zero API cost.

ccl is a thin, zero-dependency wrapper around the real claude binary, operating in an isolated profile (~/.claude-isvalorum). Your existing Claude Code config remains completely untouched.

Key Highlights

⚡ Zero API Cost

Develop unlimited code sessions with your local Apple Silicon GPU and Ollama models.

🧠 Instant VRAM Eviction

Switching models with /model immediately unloads the previous model, preventing OOM crashes.

🎛️ LM-Studio-Style Tune Panel

Interactively tune context size, GPU offloading, and temperature with ccl --tune.

🤗 Hugging Face Search

Search, preview, and download GGUFs straight from Hugging Face using /find-model.

Installation

git clone https://github.com/nsozturk/claude-code-local.git
cd claude-code-local
bash scripts/install-symlinks.sh
ccl

Quick Start

# Launch Claude Code on your default local model
ccl

# Launch with a specific local model
ccl -m qwen2.5-coder:32b

# Inside a ccl session:
/model                  # Switch models with instant VRAM eviction
/find-model qwen coder  # Search & import GGUF straight from Hugging Face

Command Reference

Command Action
ccl Start Claude Code on default local model.
ccl -m <model> Launch session with designated model.
ccl --list Show available local models and resident VRAM status.
ccl --tune [model] Open interactive context & parameter tuning panel.
ccl --fix-ctx Generate -ctx128k / -ctx256k variants automatically.
ccl --unload Immediately evict all models from VRAM.
ccl --bridge status Check health of the translation bridge daemon.

Importing Models

Import GGUF files directly from local storage, LM Studio, or Hugging Face:

# List models already in LM Studio
python3 scripts/import-model.py --list-lmstudio

# Import local GGUF without re-downloading
python3 scripts/import-model.py ~/.lmstudio/models/.../model.gguf

# Import direct from Hugging Face
python3 scripts/import-model.py hf.co/TheUser/TheRepo-GGUF:Q4_K_M

Architecture & Lifecycle Layer

Claude Code communicates with a lightweight Python bridge (bin/claude-ollama-bridge, port 11435) that translates the Anthropic Messages API to Ollama and manages VRAM memory lifecycle.

Full design measurements and memory benchmarks are documented in docs/ARCHITECTURE.md.

License

Distributed under the MIT License. Copyright © 2026 Enes Öztürk.