claude-codex-local 0.6.0

Local backend bridge for Claude Code and Codex.

Hit your limit? Need privacy? Just swap the model.

One alias. Claude Code or Codex on a local model. Skills, agents, MCP servers — all intact.

Quota hit mid-session? cc keeps you going on a local model, no context lost. Code that can't leave your machine? Everything runs offline after model download. Don't want to rewire your workflow? Your ~/.claude, skills, agents, and MCP servers carry over untouched.

Get Started → · Landing page →

---

Features

Feature	What you get
Ollama first-class	ollama launch — no duplicated config, no custom Modelfiles
Config untouched	All skills, statusline, agents, plugins, and MCP servers carry over
Smart model selection	llmfit analyses your hardware and picks the best quantization that fits (optional — wizard prompts to install only when needed)
Resume on failure	Wizard persists progress — --resume picks up from the last completed step
Idempotent aliases	Re-running the wizard replaces the existing alias block, never appends
Cloud fallback	Run claude / codex directly (no prefix) to switch back instantly

---

Quick Start

Install from PyPI (recommended)

pip install claude-codex-local

Or with uv:

uv tool install claude-codex-local

Then run the setup wizard:

ccl

One-command install (no clone required)

bash <(curl -sSL https://raw.githubusercontent.com/luongnv89/claude-codex-local/main/install.sh)

Or with wget:

bash <(wget -qO- https://raw.githubusercontent.com/luongnv89/claude-codex-local/main/install.sh)

Use bash <(...), not curl … | bash. The wizard is interactive and needs a real TTY — piping steals stdin.

Override defaults with env vars:

CCL_REF=v0.6.0 CCL_INSTALL_DIR=~/tools/claude-codex-local \
  bash <(curl -sSL https://raw.githubusercontent.com/luongnv89/claude-codex-local/main/install.sh)

Install from a clone

git clone https://github.com/luongnv89/claude-codex-local.git
cd claude-codex-local

python3 -m venv .venv && source .venv/bin/activate
pip install -e .

ccl

After setup

Reload your shell so the alias is available:

source ~/.zshrc   # or source ~/.bashrc

Then run:

cc        # Claude Code → local model
cx        # Codex CLI → local model

---

Wizard Steps

graph TD
    A[1. Discover environment] --> B[2. Install missing components]
    B --> C[3. Pick harness + engine]
    C --> D[4. Pick model]
    D --> E[5. Smoke test engine]
    E --> F[6. Wire harness]
    F --> G[7. Install helper + aliases]
    G --> H[8. Verify launch end-to-end]
    H --> I[9. Generate guide.md]

See guide.example.md for the personalized daily-use guide the wizard generates.

---

Usage

ccl                                             # run the interactive first-run wizard
ccl setup --harness claude --engine ollama      # skip the prefs picker
ccl setup --non-interactive                     # CI-friendly install
ccl setup --resume                              # resume after a failure
ccl find-model                                  # standalone model recommendation
ccl doctor                                      # wizard state + presence check
ccl --version                                   # print version and exit

Advanced / debug (no user binary — run as a Python module):

python -m claude_codex_local.core profile      # full hardware profile as JSON
python -m claude_codex_local.core recommend    # llmfit-only model recommendation
python -m claude_codex_local.core adapters     # list all engine adapters

---

Prerequisites

• macOS or Linux with zsh or bash
• Python 3.10+
• At least one harness: Claude Code or Codex CLI
• At least one engine: Ollama (recommended), LM Studio, or llama.cpp
• llmfit on PATH (optional — for automatic model selection)

---

Proven Paths

Harness	Engine	Model	Status
Claude Code	Ollama	gemma4:26b	Verified end-to-end
Codex CLI	Ollama	gemma4:26b	Verified
Codex CLI	Ollama	qwen2.5-coder:0.5b	Verified
Claude Code	LM Studio	Qwen3 family	Blocked — 400 thinking.type; wizard warns and recommends alternatives
Any	llama.cpp	any	Inline-env code path exists, no live proof yet

---

Rollback

# Remove the fenced block from ~/.zshrc (between the marker lines)
rm -rf .claude-codex-local

That's it. Your ~/.claude and ~/.codex are unchanged.

---

Architecture details

Three layers

1. Machine profile + model recommendation (claude_codex_local/core.py) — dumps a JSON snapshot of installed harnesses/engines/llmfit/disk, runs llmfit for ranked model recommendations, and provides a doctor command for pretty-printing wizard state.

2. Interactive wizard (claude_codex_local/wizard.py) — 9 steps from discovery to ready-to-use daily alias. Persists progress in .claude-codex-local/wizard-state.json so --resume picks up after a failure.

3. Helper scripts + shell aliases — .claude-codex-local/bin/cc (or cx) is a short bash wrapper. For Ollama it runs ollama launch claude|codex --model <tag>. For LM Studio / llama.cpp it sets inline env vars and execs the real harness. A fenced block in ~/.zshrc / ~/.bashrc declares the aliases.

Why ollama launch

ollama launch claude --model <tag> is an official Ollama subcommand that sets the right env vars internally and execs the user's real claude binary against the local daemon — using ~/.claude as-is.

This means:

• No duplicated ~/.claude directory
• No custom Modelfile or ollama create
• No ANTHROPIC_CUSTOM_MODEL_OPTION to manage manually
• cc just works

Claude Code → LM Studio / llama.cpp env vars

Env var	LM Studio	llama.cpp
ANTHROPIC_BASE_URL	http://localhost:1234	http://localhost:8001
ANTHROPIC_API_KEY	lmstudio	sk-local
ANTHROPIC_CUSTOM_MODEL_OPTION	<tag>	<tag>
ANTHROPIC_CUSTOM_MODEL_OPTION_NAME	Local (lmstudio) <tag>	Local (llamacpp) <tag>
CLAUDE_CODE_ATTRIBUTION_HEADER	"0"	"0"
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC	"1"	"1"

Codex CLI → Ollama

ollama launch codex --model <tag> -- --oss --local-provider=ollama

The --oss --local-provider=ollama flags are required after -- because Codex otherwise tries to route through the ChatGPT account and rejects non-OpenAI model names.

Qwen3 + Claude Code

Claude Code sends a thinking payload that Qwen3 reasoning models interpret as an unterminated <think> block. The wizard detects Qwen3 model names at pick time and recommends Gemma 3 or Qwen 2.5 Coder instead.

Project structure

.
├── claude_codex_local/
│   ├── __init__.py             # Package metadata + __version__
│   ├── wizard.py               # Interactive setup wizard + `ccl` CLI
│   └── core.py                 # Machine profile, engine adapters, llmfit bindings
├── scripts/
│   └── e2e_smoke.sh            # End-to-end smoke test
├── docs/
│   ├── poc-wizard.md           # 9-step wizard architecture
│   ├── poc-architecture.md     # System design overview
│   ├── poc-bootstrap.md        # Bootstrap / install flow
│   └── poc-proof.md            # Design rationale
├── tests/                      # pytest test suite
├── install.sh                  # One-command remote installer
└── pyproject.toml              # Project metadata and tool config

Tech stack

Layer	Tool
Language	Python 3.10+
UI / prompts	questionary, rich
Linting	ruff
Type checking	mypy
Testing	pytest + pytest-cov
Security	bandit, detect-secrets
Pre-commit	pre-commit

Local state

Everything written by the bridge goes under .claude-codex-local/. Override with CLAUDE_CODEX_LOCAL_STATE_DIR.

Contributing

Contributions are welcome. Read CONTRIBUTING.md before opening a PR.

For security issues, see SECURITY.md.

---

MIT — © 2024 Luong NGUYEN