<div align="center">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="assets/logo/logo-white.svg">
<source media="(prefers-color-scheme: light)" srcset="assets/logo/logo-black.svg">
<img src="assets/logo/logo-full.svg" alt="CCL โ Claude Codex Local" height="60">
</picture>
[](LICENSE)
[](https://www.python.org/downloads/)
[](https://github.com/luongnv89/claude-codex-local/actions/workflows/ci.yml)
[](https://github.com/astral-sh/ruff)
[](https://pypi.org/project/claude-codex-local/)
[](https://pypi.org/project/claude-codex-local/)
[](https://pepy.tech/project/claude-codex-local)
</div>
# Hit your limit? Need privacy? Just swap the model.
One alias. Claude Code, Codex, or [Pi](https://pi.dev/) on a local model. Skills, agents, MCP servers โ 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 โ**](#quick-start) ยท [**Landing page โ**](https://luongnv89.github.io/claude-codex-local/)
---
## Features
| Feature | What you get |
| --------------------- | --------------------------------------------------------------------------------------------------------------------------------- |
| Ollama first-class | `ollama launch` โ 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 (lazy hardware scan; pass `--run-llmfit` to refresh) |
| Resume on failure | Wizard persists progress โ `--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` / `pi` directly (no prefix) to switch back instantly |
---
## Quick Start
### Install from PyPI (recommended)
```bash
pip install claude-codex-local
```
Or with uv:
```bash
uv tool install claude-codex-local
```
Then run the setup wizard:
```bash
ccl
```
### One-command install (no clone required)
```bash
bash <(curl -sSL https://raw.githubusercontent.com/luongnv89/ccl/main/install.sh)
```
Or with wget:
```bash
bash <(wget -qO- https://raw.githubusercontent.com/luongnv89/ccl/main/install.sh)
```
> Use `bash <(...)`, not `curl โฆ | bash`. The wizard is interactive and needs a real TTY โ piping steals stdin.
Override defaults with env vars:
```bash
CCL_REF=v0.14.0 CCL_INSTALL_DIR=~/tools/claude-codex-local \
bash <(curl -sSL https://raw.githubusercontent.com/luongnv89/ccl/main/install.sh)
```
### Install from a clone
```bash
git clone https://github.com/luongnv89/claude-codex-local.git
cd claude-codex-local
```
```bash
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
```
```bash
ccl
```
### After setup
Reload your shell so the alias is available:
```bash
source ~/.zshrc # or source ~/.bashrc
```
Then run:
```bash
cc # Claude Code โ local model
cx # Codex CLI โ local model
ccp # Pi โ local model
```
---
## Wizard Steps
```mermaid
graph TD
A[1. Discover environment<br>cached snapshot; lazy hardware scan] --> B[2. Defer install prompts]
B --> C[3. Pick harness + engine<br>selected choices are checked live]
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`](guide.example.md) for the personalized daily-use guide the wizard generates.
---
## Usage
```bash
ccl # run the interactive first-run wizard
ccl setup --harness claude --engine ollama # skip the prefs picker
ccl setup --harness pi --engine ollama # Pi (https://pi.dev/) + local model
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 run # launch the configured session interactively
ccl run -p "what is 2+2?" # one-shot: drive CCL from another agent
ccl run --native-params -- --dangerously-skip-permissions
# forward harness-native flags verbatim
ccl --version # print version and exit
```
`ccl run -p "<prompt>"` runs the harness in non-interactive mode (Claude
Code's `-p`, Codex's `exec`, Pi's `--print`) so external agents and CI scripts
can drive a local model end-to-end without keystrokes. Without `-p`, behavior
matches the `cc` / `cx` / `ccp` alias and the session starts interactively.
### Forwarding harness-native flags (`--native-params`)
`ccl run --native-params -- <ARGSโฆ>` forwards everything after `--native-params`
verbatim to the launched harness. It is a generic escape hatch for harness
options that ccl does not wrap explicitly (e.g. Claude Code's
`--dangerously-skip-permissions`). Must be the last flag on the line; use `--`
to make the boundary between ccl flags and native ones explicit. ccl does not
validate native params โ the harness does.
```bash
# Claude Code: skip permission prompts (interactive)
ccl run --native-params -- --dangerously-skip-permissions
# Codex: set extra exec-time flag (one-shot)
ccl run -p "summarize this repo" --native-params -- --some-codex-flag
# Pi: forward any pi-native option
ccl run --native-params -- --some-pi-option value
```
Advanced / debug (no user binary โ run as a Python module):
```bash
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
python -m claude_codex_local.core engine ollama config
python -m claude_codex_local.core engine vllm benchmark --model <served-model>
```
---
## Engine Lifecycle Scripts
Engine-specific lifecycle behavior lives under `claude_codex_local/engines/`.
Each supported engine owns the same script contract:
| Action | Purpose |
|---|---|
| `install` | Installation command or manual setup instructions |
| `config` | Endpoint, key-file, and environment settings |
| `optimize` | Engine-specific tuning recommendations |
| `test` | Smoke test for the selected model or endpoint |
| `benchmark` | Lightweight timing/throughput check where safe |
Supported engine packages are `ollama`, `lmstudio`, `llamacpp`, `vllm`,
`router9` (engine name `9router`), and `openrouter`. Power users can customize
one engine by editing only that engine package. Typical callers should use the
uniform dispatcher:
```bash
python -m claude_codex_local.core engine <engine> <action> [--model <tag>] [--execute]
```
`test` and `benchmark` dry-run unless `--execute` is passed. Adding another
engine means adding a package with those five modules and `ENGINE_NAME` metadata;
the dispatcher discovers it without new core branching.
---
## Sharing Context Between Agents 
`ccl run` automatically bridges conversation context across local harnesses
so you can hand off mid-task between Claude Code, Codex, and Pi without
re-explaining what you were doing. The bridge runs in two halves:
- **Post-run capture (both paths)** โ after the harness exits, CCL reads its
native session file for the current `$PWD`
(`~/.claude/projects/...`, `~/.codex/sessions/...`, `~/.pi/agent/sessions/...`)
and imports the cleaned, redacted messages into
`~/.claude-codex-local/sessions/<harness>.jsonl`.
- **Pre-run injection (one-shot only)** โ when you `ccl run -p PROMPT`, the
freshest *other* harness's transcript for `$PWD` is rendered as a
`[prior context, agent=โฆ]\nโฆ\n[end prior context]` block and prepended
to `PROMPT`. Interactive sessions don't inject โ each harness has its
own `--resume`/`--continue` and stdin is the user's TTY. The
cc-interactive โ cx-one-shot handoff is the prototypical flow.
```bash
ccl run -p "continue from where claude left off" # auto-injects context if claude has one
ccl run --no-context -p "fresh start" # disable bridge (also CCL_SESSION_BRIDGE=0)
ccl session list # show captured per-harness JSONL files
ccl session show claude # print one harness's messages (JSON)
ccl session sync --from claude --to codex # manual copy if you want to force it
ccl session truncate codex --keep 50 # keep only the last N (--keep required)
ccl session clear codex # delete one harness's file
```
**Scope guards** keep the bridge predictable:
- **cwd-scoped** โ context only flows for the same repository. Switching
directories starts fresh.
- **7-day staleness cap** โ native sessions older than a week are ignored
so months-old transcripts can't get silently re-injected. The injection
banner shows the source's age (`last activity 6m ago`) so you can spot a
stale handoff before it ships to the model.
- **Source picker** โ when multiple non-self harnesses have history, the
most recently modified file wins. Same-harness one-shot continuity
(`cc -p "foo"` then `cc -p "bar"`) is *not* covered โ use Claude Code's
own `--resume` for that.
- **Boilerplate filter** โ `AGENTS.md` / `CLAUDE.md` re-dumps, slash-command
echoes, skill loads, tool calls, reasoning traces, and other harness
internals are dropped on import; only user/assistant text survives.
- **Redaction** โ best-effort scrub of common token shapes (OpenAI,
Anthropic, AWS, GitHub PAT/OAuth, Slack, GitLab, Google API) on every
import and sync. Treat the JSONL files as semi-sensitive โ the scrub
covers known patterns, not arbitrary secrets in prose.
- **Idempotent** โ re-running `ccl run` against the same native file does
not duplicate messages; a content-hash dedup key skips already-imported
rows.
The capture path is exercised on both interactive and one-shot via the same
post-run import. The injection path is currently wired only on the one-shot
(`-p`) branch.
State directory can be overridden with `CLAUDE_CODEX_LOCAL_STATE_DIR`; the
native-home base can be overridden with `CCL_NATIVE_HOME_OVERRIDE` (useful
for tests and CI). `--keep` is **required** on `truncate` to prevent
accidental wipes; use `ccl session clear` if you want to remove the file
entirely.
---
## Prerequisites
- macOS or Linux with zsh or bash
- Python 3.11+
- At least one harness: [Claude Code](https://claude.ai/code), [Codex CLI](https://github.com/openai/codex), or [Pi](https://pi.dev/) (`npm install -g @earendil-works/pi-coding-agent`) โ Pi is the model-agnostic terminal coding harness whose tagline is โThere are many agent harnesses, but this one is yours.โ
- At least one engine: [Ollama](https://ollama.com) (recommended), [LM Studio](https://lmstudio.ai), [vLLM](https://github.com/vllm-project/vllm), llama.cpp, [9router](https://github.com/decolua/9router) (local cloud-routing proxy), or [OpenRouter](https://openrouter.ai) (hosted SaaS)
- [`llmfit`](https://github.com/luongnv89/llmfit) on `PATH` (optional โ 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 |
| Pi | Ollama | any local tag | Supported via isolated Pi `models.json` and `ccp` alias |
| Claude Code | LM Studio | Qwen3 family | Blocked โ `400 thinking.type`; wizard warns and recommends alternatives |
| Any | llama.cpp | any | Inline-env code path exists, no live proof yet |
| Any | vLLM | any | New in 0.8.0 โ adapter shipped with tests |
| Claude Code | 9router | `kr/claude-sonnet-4.5` | New in 0.9.0 โ cloud-routed via `cc9` alias; existing `cc` is untouched |
| Codex CLI | 9router | `kr/claude-sonnet-4.5` | New in 0.9.0 โ cloud-routed via `cx9` alias; existing `cx` is untouched |
| Pi | 9router | `kr/claude-sonnet-4.5` | Cloud-routed via `cp9`; existing `ccp` is untouched |
| Claude Code | OpenRouter | `anthropic/claude-sonnet-4.6` | Hosted SaaS via `cco` alias; existing `cc` is untouched (#83) |
| Codex CLI | OpenRouter | `anthropic/claude-sonnet-4.6` | Hosted SaaS via `cxo` alias; existing `cx` is untouched (#83) |
| Pi | OpenRouter | `anthropic/claude-sonnet-4.6` | Hosted SaaS via `cpo`; existing `ccp` is untouched (#83) |
---
## Remote engine endpoints
`ccl` can consume Ollama, llama.cpp, and vLLM servers running on another
machine. The wizard probes the endpoint over HTTP and does **not** require the
remote engine binary to be installed locally.
### Interactive flow (primary)
When you run `ccl setup` and pick `ollama`, `llamacpp`, or `vllm` as your
engine, step 3 follows up with:
1. **Local or remote?** โ prompt: *Run `<engine>` locally, or use a remote
endpoint?* Default is **Local**. Pick **Remote** to point at another host.
2. **Base URL** โ prompt: *Remote `<engine>` base URL (scheme + host + port,
no path):* The wizard validates the input must be a bare base URL. Paths
(e.g. `http://gpu-box.local:8001/v1`), queries, and fragments are rejected;
the engine probes add the correct suffix per engine (`/api/tags` for Ollama,
`/health` and `/v1/models` for llama.cpp, `/v1/models` for vLLM).
3. **API key** โ for `llamacpp` and `vllm` only, a masked password prompt:
*`<engine>` API key (leave empty for no auth):*. Ollama does not prompt.
4. **Persist to shell rc?** โ prompt: *Also persist these env vars to your
shell rc?* Default **No**. Pick Yes to write a fenced
`# >>> claude-codex-local:remote:<engine> >>>` block into `~/.zshrc` /
`~/.bashrc` so future shells inherit the same endpoint.
**Selecting Remote skips the local-binary install step** โ the wizard re-probes
the URL you just provided and proceeds to model selection without offering to
install Ollama / llama.cpp / vLLM on the local machine.
Cancel any prompt with `Ctrl-C` / `Esc` to fall back to the local-install path.
### Non-interactive / CI
Set the env vars **before** running `ccl setup` (or pass `--non-interactive`).
The wizard picks them up via `core.py` and treats the engine as remote without
asking:
```bash
# Ollama native API and OpenAI-compatible /v1 endpoint
export OLLAMA_HOST=http://gpu-box.local:11434
# llama.cpp server (OpenAI-compatible API under /v1)
export LLAMACPP_BASE_URL=http://gpu-box.local:8001
# Optional if your reverse proxy / server requires bearer auth:
export LLAMACPP_API_KEY=...
# vLLM OpenAI-compatible server
export VLLM_BASE_URL=http://gpu-box.local:8000
export VLLM_API_KEY=... # optional; vLLM only checks this if configured
```
Each URL must be the **base** of the engine host โ scheme, host, and port only.
Do **not** append `/v1`, `/api`, or any other path. A trailing path silently
double-suffixes to a 404; the wizard warns and strips it, but the right input
is the bare host.
Local and remote engines can coexist: unset the relevant env var (or remove
the fenced rc block) to go back to the localhost default, or choose a different
engine in the wizard.
---
## 9router quick-start
[9router](https://github.com/decolua/9router) is a local proxy that exposes an OpenAI-compatible API on `http://localhost:20128/v1` and routes calls to cloud models such as `kr/claude-sonnet-4.5`. Picking 9router as the engine adds a **new** `cc9` (Claude), `cx9` (Codex), or `cp9` (Pi) alias and leaves your existing `cc` / `cx` / `ccp` aliases untouched.
### Installing and running 9router
**Step 1: Install 9router**
```bash
# Using npm (recommended)
npm install -g 9router
# Or using yarn
yarn global add 9router
# Or using pnpm
pnpm add -g 9router
```
**Step 2: Get your API key**
1. Visit the [9router dashboard](https://9router.com/dashboard) and sign up or log in
2. Navigate to API Keys section
3. Create a new API key and copy it
**Step 3: Start the 9router service**
```bash
# Start 9router with your API key
9router start --api-key YOUR_API_KEY_HERE
# Or set it as an environment variable
export ROUTER9_API_KEY=YOUR_API_KEY_HERE
9router start
# The service will start on http://localhost:20128
```
**Step 4: Verify 9router is running**
```bash
# Check if the service is responding
curl http://localhost:20128/v1/models
# You should see a list of available models
```
**Step 5: Configure CCL to use 9router**
```bash
# Interactive setup (wizard will prompt for API key, then list models from /v1/models)
ccl setup --engine 9router
# Non-interactive (CI / scripted):
CCL_9ROUTER_API_KEY=<paste-here> CCL_9ROUTER_MODEL=kr/claude-sonnet-4.5 \
ccl setup --engine 9router --harness claude --non-interactive
```
### How the wizard configures 9router
The wizard:
1. Asks for the 9router API key and writes it to `~/.claude-codex-local/9router-api-key` with `chmod 0600`. The helper script reads this file at exec time via `$(cat โฆ)` โ the key is never embedded in the script body or wizard state file.
2. Fetches available models via `GET /v1/models` and shows them as a selectable list. If the endpoint is unreachable or returns no models, the wizard falls back to manual model-name entry.
3. Verifies reachability via `GET /v1/models` only. **It deliberately does not call `/chat/completions`** during smoke-test or verify, because 9router routes to paid cloud models. The verification record is `{"ok": true, "via": "9router-models-endpoint", "skipped_chat": true}`.
4. Installs `cc9` (or `cx9`) into your shell rc as a new fenced block (`# >>> claude-codex-local:claude9 >>>`), leaving any existing `cc` / `cx` block alone.
**Tip:** `cc9` and `cc` can coexist on the same machine โ pick `cc9` when you want to burn cloud quota for a tough prompt, and `cc` (Ollama / LM Studio / llama.cpp) for everyday work.
### Claude Code โ 9router env vars
| Env var | 9router |
| ------------------------------------------ | ------------------------------------------------------------- |
| `ANTHROPIC_BASE_URL` | `http://localhost:20128/v1` |
| `ANTHROPIC_AUTH_TOKEN` | `$(cat ~/.claude-codex-local/9router-api-key)` (read at exec) |
| `ANTHROPIC_API_KEY` | `$(cat ~/.claude-codex-local/9router-api-key)` (read at exec) |
| `ANTHROPIC_CUSTOM_MODEL_OPTION` | `<tag>` (e.g. `kr/claude-sonnet-4.5`) |
| `ANTHROPIC_CUSTOM_MODEL_OPTION_NAME` | `9router <tag>` |
| `CLAUDE_CODE_ATTRIBUTION_HEADER` | `"0"` |
| `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` | `"1"` |
For Codex: `OPENAI_BASE_URL=http://localhost:20128/v1`, `OPENAI_API_KEY=$(cat โฆ)`.
---
## OpenRouter quick-start
[OpenRouter](https://openrouter.ai) is a hosted-SaaS cloud-routing service that exposes an OpenAI-compatible API at `https://openrouter.ai/api/v1` and forwards calls to dozens of cloud models (Claude, GPT-4o, Llama 3.1, Mistral, and many more). Unlike 9router, there is **no daemon to install** โ only an API key. Picking OpenRouter as the engine adds a **new** `cco` (Claude), `cxo` (Codex), or `cpo` (Pi) alias and leaves your existing `cc` / `cx` / `ccp` aliases untouched.
### Setting up OpenRouter
**Step 1: Get your API key**
1. Visit [openrouter.ai/keys](https://openrouter.ai/keys) and sign up or log in
2. Create a new API key and copy it
3. Optionally fund your account or set a per-key spending limit
**Step 2: Configure CCL to use OpenRouter**
```bash
# Interactive setup (wizard will prompt for API key)
ccl setup --engine openrouter
# Non-interactive (CI / scripted):
CCL_OPENROUTER_API_KEY=<paste-here> CCL_OPENROUTER_MODEL=anthropic/claude-sonnet-4.6 \
ccl setup --engine openrouter --harness claude --non-interactive
```
### How the wizard configures OpenRouter
The wizard:
1. Asks for the OpenRouter API key and writes it to `~/.claude-codex-local/openrouter-api-key` with `chmod 0600`. The helper script reads this file at exec time via `$(cat โฆ)` โ the key is never embedded in the script body or wizard state file.
2. Verifies reachability via `GET /models` only. **It deliberately does not call `/chat/completions`** during smoke-test or verify, because OpenRouter routes to paid cloud models. The verification record is `{"ok": true, "via": "openrouter-models-endpoint", "skipped_chat": true}`.
3. Installs `cco` (or `cxo` / `cpo`) into your shell rc as a new fenced block (`# >>> claude-codex-local:claudeo >>>`), leaving any existing `cc` / `cx` / `ccp` block alone.
**Tip:** `cco` and `cc` can coexist on the same machine โ pick `cco` when you want a hosted model, and `cc` (Ollama / LM Studio / llama.cpp) for everyday local work.
### Claude Code โ OpenRouter env vars
| Env var | OpenRouter |
| ------------------------------------------ | ------------------------------------------------------------------- |
| `ANTHROPIC_BASE_URL` | `https://openrouter.ai/api/v1` |
| `ANTHROPIC_AUTH_TOKEN` | `$(cat ~/.claude-codex-local/openrouter-api-key)` (read at exec) |
| `ANTHROPIC_API_KEY` | `$(cat ~/.claude-codex-local/openrouter-api-key)` (read at exec) |
| `ANTHROPIC_CUSTOM_MODEL_OPTION` | `<tag>` (e.g. `anthropic/claude-sonnet-4.6`) |
| `ANTHROPIC_CUSTOM_MODEL_OPTION_NAME` | `OpenRouter <tag>` |
| `HTTP_REFERER` | `https://github.com/luongnv89/ccl` (OpenRouter attribution header) |
| `X_TITLE` | `claude-codex-local` (OpenRouter attribution header) |
| `CLAUDE_CODE_ATTRIBUTION_HEADER` | `"0"` |
| `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` | `"1"` |
For Codex: `OPENAI_BASE_URL=https://openrouter.ai/api/v1`, `OPENAI_API_KEY=$(cat โฆ)`.
Override the base URL via `CCL_OPENROUTER_BASE_URL` and the model via `CCL_OPENROUTER_MODEL`.
---
## Rollback
```bash
# Remove the fenced block from ~/.zshrc (between the marker lines)
rm -rf .claude-codex-local
```
Each fence block (`claude` / `codex` / `claude9` / `codex9` / `claudeo` / `codexo`) is independent โ you can remove just one without touching the others. Your `~/.claude` and `~/.codex` are unchanged.
---
<details>
<summary>Architecture details</summary>
### Three layers
1. **Machine profile + model recommendation** (`claude_codex_local/core.py`) โ 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`) โ 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** โ `.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 โ 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 โ 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"` |
#### MTP (Multi-Token Prediction) models
llama.cpp โฅ 2026-05-16 supports Multi-Token Prediction speculative decoding for models like [`unsloth/Qwen3.6-27B-MTP-GGUF`](https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF), giving ~1.5โ2ร faster inference at no accuracy cost. CCL detects MTP variants automatically (GGUF metadata probe, then `*mtp*` filename match) and adds the required `--spec-type draft-mtp --spec-draft-n-max 5` flags to the auto-started `llama-server`. Override via env:
| Env var | Effect |
| ----------------------------- | ------------------------------------------------------------------------------------- |
| `LLAMACPP_MTP_ENABLED` | `0` forces MTP off; `1` forces it on. Unset โ auto-detect from GGUF / filename. |
| `LLAMACPP_SPEC_DRAFT_N_MAX` | Override `--spec-draft-n-max` (default `5`; valid 1โ16; Unsloth recommends 3โ6). |
Both env vars are read at `claude_codex_local` import time, so set them before invoking the CLI (in your shell, in a wrapper script, or via `env LLAMACPP_MTP_ENABLED=0 ccl โฆ`). Mutating `os.environ` after the package has loaded has no effect.
Note: llama.cpp does not yet support combining `--spec-type draft-mtp` with `--mmproj` or `-np`/`--parallel > 1`. CCL's auto-started `llama-server` does not pass those flags today, so this is a forward-looking caveat: if you run `llama-server` manually alongside those flags, set `LLAMACPP_MTP_ENABLED=0`.
### Codex CLI โ Ollama
```bash
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.
</details>
<details>
<summary>Project structure</summary>
```
.
โโโ claude_codex_local/
โ โโโ __init__.py # Package metadata + __version__
โ โโโ engines/ # Per-engine install/config/optimize/test/benchmark scripts
โ โโโ wizard.py # Interactive setup wizard + `ccl` CLI
โ โโโ core.py # Machine profile, engine adapters, llmfit bindings
โโโ scripts/
โ โโโ e2e_smoke.sh # End-to-end smoke test
โโโ docs/
โ โโโ poc-wizard.md # 9-step wizard architecture
โ โโโ poc-architecture.md # System design overview
โ โโโ poc-bootstrap.md # Bootstrap / install flow
โ โโโ poc-proof.md # Design rationale
โโโ tests/ # pytest test suite
โโโ install.sh # One-command remote installer
โโโ pyproject.toml # Project metadata and tool config
```
</details>
<details>
<summary>Tech stack</summary>
| Layer | Tool |
| ------------- | --------------------------------------------------------------------------------------------------- |
| Language | Python 3.11+ |
| UI / prompts | [questionary](https://github.com/tmbo/questionary), [rich](https://github.com/Textualize/rich) |
| Linting | [ruff](https://github.com/astral-sh/ruff) |
| Type checking | [mypy](https://mypy-lang.org) |
| Testing | [pytest](https://pytest.org) + pytest-cov |
| Security | [bandit](https://github.com/PyCQA/bandit), [detect-secrets](https://github.com/Yelp/detect-secrets) |
| Pre-commit | [pre-commit](https://pre-commit.com) |
</details>
<details>
<summary>Local state</summary>
Everything written by the bridge goes under `.claude-codex-local/`. Override with `CLAUDE_CODEX_LOCAL_STATE_DIR`.
</details>
<details>
<summary>Contributing</summary>
Contributions are welcome. Read [CONTRIBUTING.md](CONTRIBUTING.md) before opening a PR.
For security issues, see [SECURITY.md](SECURITY.md).
</details>
---
[MIT](LICENSE) โ ยฉ 2026 Luong NGUYEN