Skip to content

README.md

Latest commit

 

History

History
107 lines (72 loc) · 4.06 KB

README.md

File metadata and controls

107 lines (72 loc) · 4.06 KB
Raw

Claw Code Local

Claude Code experience. Local models. Zero API costs.

Run a full-featured coding agent CLI against Ollama, LM Studio, or any OpenAI-compatible endpoint — entirely on your machine, with your own models.

Built on top of claw-code-parity, a clean-room Rust reimplementation of the Claude Code harness.

What this is

Claw Code Local patches the claw CLI to use any LLM provider, not just Anthropic. The upstream project built a full multi-provider API layer but never wired it into the actual binary — it was hardcoded to require Anthropic credentials. We fixed that, along with a streaming render bug that caused terminal output to smash sections together.

What you get:

  • Interactive REPL with session persistence, slash commands, markdown rendering
  • Tool use: file read/write/edit, grep, glob, bash execution, git integration
  • /commit, /diff, /pr, /review, /mcp, /agents, /skills and 130+ slash commands
  • Works with any model your hardware can run
  • Sessions auto-save and can be resumed

Quickstart

Prerequisites

  • Rust toolchain (1.70+)
  • Ollama running locally (or any OpenAI-compatible endpoint)
  • A model pulled in Ollama (e.g. ollama pull qwen3:14b)

Build

git clone https://github.com/codetwentyfive/claw-code-local.git
cd claw-code-local/rust
cargo build -p rusty-claude-cli --release

The binary is at rust/target/release/claw.exe (Windows) or rust/target/release/claw (Linux/macOS).

Configure

Set two environment variables. The API key value doesn't matter for Ollama — it just needs to be non-empty.

Bash / Zsh:

export OPENAI_API_KEY=ollama
export OPENAI_BASE_URL=http://localhost:11434/v1

PowerShell (add to $PROFILE):

$env:OPENAI_API_KEY = "ollama"
$env:OPENAI_BASE_URL = "http://localhost:11434/v1"

Optionally add the binary to your PATH for global access.

Run

# Interactive REPL
claw --model qwen3:14b

# One-shot prompt
claw --model qwen3:14b "explain this codebase"

# Switch models mid-session with /model
/model qwen3.5-35b-uncensored:iq3m

# Resume your last session
claw --resume latest

Supported providers

The CLI auto-detects the provider from the model name and environment variables:

Provider Models Env vars
Ollama Any model you've pulled OPENAI_API_KEY=ollama OPENAI_BASE_URL=http://localhost:11434/v1
LM Studio Any loaded model OPENAI_API_KEY=lmstudio OPENAI_BASE_URL=http://localhost:1234/v1
OpenAI gpt-4o, o1, etc. OPENAI_API_KEY=sk-...
xAI (Grok) grok-3, grok-3-mini XAI_API_KEY=xai-...
Anthropic claude-opus, sonnet, haiku ANTHROPIC_API_KEY=sk-ant-...
Any OpenAI-compatible Depends on provider OPENAI_API_KEY=... OPENAI_BASE_URL=https://...

What we changed from upstream

Two patches, both original:

  1. Multi-provider CLI wiring — The api crate already had ProviderClient with OpenAI-compat and xAI support, but the CLI hardcoded AnthropicClient and always required Anthropic credentials. We swapped it to ProviderClient::from_model_with_anthropic_auth() so the provider is auto-detected from the model name and env vars.

  2. Streaming markdown render fix — render_markdown() called trim_end() which stripped trailing newlines. In streaming mode each chunk is rendered independently, so block separators (between tables, headings, paragraphs) got eaten and everything ran together on one line. Added a streaming-safe render path that preserves block spacing.

Upstream

This is a fork of ultraworkers/claw-code-parity, which is itself a clean-room Rust reimplementation of Claude Code's agent harness — not a copy of Anthropic's source code. See the upstream README and PARITY.md for the full porting status.

To sync with upstream:

git fetch upstream
git merge upstream/main

License

MIT — same as upstream.