Questions and answers
Answers sourced from the code, docs, and release notes. Sources are cited under each answer.
ClawCodex is a terminal-native coding agent and a production-oriented Python rebuild of Claude Code. It runs a real tool-calling loop, streams replies in an inline REPL, edits local workspaces behind permission gates, and puts every major LLM provider behind one runtime.
Sources: README.md docs/ARCHITECTURE.md
Claude Code targets Claude models only and is proprietary. ClawCodex is open source (MIT), written in idiomatic Python, and built around multi-provider support: Anthropic, OpenAI, Z.ai GLM, MiniMax, OpenRouter, and DeepSeek behind the same agent loop. It is an independent educational project and is not affiliated with Anthropic.
Sources: README.md
Clone the repo, create a Python 3.10+ virtualenv, and install requirements:
git clone https://github.com/agentforce314/clawcodex.git cd clawcodex python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt
Then run python -m src.cli login to configure a provider, and clawcodex to start. uv users can uv pip install -e ".[dev]" for the entry point. See the install page for every path.
Sources: README.md
25 providers ship today. Seven are native integrations — DeepSeek, Anthropic, OpenAI, Google Gemini, Z.ai GLM, MiniMax, and OpenRouter — alongside an OpenAI-compatible gateway registry (NVIDIA NIM, Together, Fireworks, Novita, SiliconFlow, Moonshot/Kimi, DeepInfra, Hugging Face, Volcengine, StepFun, Arcee, AtlasCloud, Xiaomi MiMo, Wanjie Ark) and three local servers that need no key (Ollama, vLLM, SGLang). Any other OpenAI-compatible endpoint also works through a custom base_url. Set the provider and key in ~/.clawcodex/config.json or run clawcodex login.
Sources: README.md docs/CONFIGURATION.md
Provider keys live in ~/.clawcodex/config.json under providers.<name>.api_key. The easiest way to write them is clawcodex login, which prompts for a provider and key; you can also edit the file directly.
Secrets that need to reach the environment — like TAVILY_API_KEY for WebSearch — go in the config env block. A matching variable already exported in your shell takes precedence over the stored value, which is handy for temporary overrides and CI.
Sources: README.md src/secret_store.py
Plan is read-only investigation — it can read files, grep, and fetch URLs but not write or run shell. Editing modes (acceptEdits, dontAsk) allow multi-step tool calls with approval. --dangerously-skip-permissions bypasses every check and is meant only for sandboxed containers with no internet. The REPL, TUI, and headless -p mode all honor the same gates.
Sources: README.md docs/MODES.md
The biggest lever is DeepSeek prefix-cache exploitation. ClawCodex keeps the request prefix byte-stable across turns, so DeepSeek's automatic prompt cache covers the entire system + tools + history span. Per-request-volatile sections (env, the mutable MEMORY.md body, plan mode) move to a trailing <system-reminder> after the conversation history, so the cached prefix never breaks. Cache-hit input is billed at only 10% of the normal rate — about $0.0435 per 1M tokens, over 200× cheaper than Claude Fable 5 — and DeepSeek-V4-Pro is already about 23× cheaper on input and 57× cheaper on output. It is gated to the deepseek provider; every other provider's request is byte-for-byte unchanged.
On top of that, /advisor mode pairs a cheap worker model with an expensive reviewer consulted only at decision points, roughly 6× cheaper than the expensive model alone. Live prompt-cache hit-rate and USD cost show in /cost and the status bar.
Sources: README.md docs/NEWS.md
An optional token-efficient setup: pair a cheap worker model with an expensive reviewer that is consulted only at decision points. On typical sessions it runs roughly 6x cheaper than using the expensive model alone, with live worker/advisor token and USD cost in the status bar. It is off by default — turn it on with /advisor <provider>:<model>.
Sources: README.md
Yes. Point the OpenAI-compatible provider at a custom base_url for vLLM, SGLang, Ollama, llama.cpp, or text-generation-webui. Local HTTP endpoints are supported for development workflows.
Sources: docs/CONFIGURATION.md
Skills are markdown SKILL.md files stored under .clawcodex/skills/<name>/ (project) or ~/.clawcodex/skills/ (user). Frontmatter sets a description, allowed-tools, and named arguments; the body is the prompt. Invoke them in the REPL as /<skill-name> <args>.
Sources: README.md
ClawCodex runs entirely on your machine. There is no telemetry and no cloud processing of your code. Permission modes are configurable per session, API keys are obfuscated in config, and .env files are ignored by default. It is built for local development workflows.
Sources: README.md SECURITY.md
On the full SWE-bench Verified split (499 instances), both agents driven by Gemini 2.5 Pro under one standardized harness, clawcodex resolved 291/499 (58.2%) versus openclaude's 265/499 (53.0%). Reproduce it locally — see eval/README.md for the workflow.
Sources: README.md eval/README.md
It is an editable Git install, so pull and reinstall: git pull, then pip install -r requirements.txt (or uv pip install -e ".[dev]") to pick up new dependencies. New features land weekly.
Sources: README.md
No CLA. Fork, branch with conventional commits (feat:, fix:, …), run python -m pytest tests/ -v, and open a PR that explains the why. Start with issues labeled good first issue. See the contribute page and CONTRIBUTING.md.
Sources: CONTRIBUTING.md