ClawSharp

ClawSharp is a C# port of Claude Code, Anthropic’s terminal-based AI assistant.

I built it because I love .NET, and there still are not many open-source code-agent projects in the .NET ecosystem. I also wanted to learn how a terminal-based coding agent is structured end to end — conversation loop, tool execution, session persistence, provider integration, and terminal UX — by building one myself.

The goal with ClawSharp is to stay close to the original Claude Code workflow while shipping as a single self-contained .NET binary that runs across Windows, Linux, and macOS without requiring Node.js at runtime.

This started as a vibe-coded side project, but I’m treating it as a real tool and a real learning project. Contributions, bug reports, and feedback are very welcome.

Why use ClawSharp?

• Single binary — no Node.js runtime after install
• Cross-platform — Windows, Linux, and macOS on x64 and ARM64
• Multi-provider — Anthropic, Gemini, OpenAI, Codex, GitHub Models, Bedrock, Vertex, Foundry, and Ollama
• Persistent sessions — resume chats, keep transcripts, and retain file history
• Extensible — supports plugins, hooks, agents, skills, and MCP servers

Current status

ClawSharp is usable today, but some parts of Claude Code are still being aligned.

Related projects

A few open projects in a similar space that I’ve looked at for reference and inspiration:

• claw-code — a public Rust implementation of the claw CLI agent harness
• openclaude — an open-source coding-agent CLI for cloud and local model providers

Install

npm install -g clawsharp
clawsharp --help

Note: Node.js is only needed for the npm-based install; GitHub Releases provide standalone binaries.

Or download a self-contained binary directly from GitHub Releases.

Update

npm update -g clawsharp
# or pin to latest:
npm install -g clawsharp@latest

Quick Start: Claude API Key

The simplest way to test ClawSharp is with a direct Anthropic API key.

Using an environment variable

# macOS / Linux
export ANTHROPIC_API_KEY="sk-ant-..."
clawsharp

# Windows (PowerShell)
$env:ANTHROPIC_API_KEY = "sk-ant-..."
clawsharp

Using CLI flags

clawsharp --provider anthropic --model claude-sonnet-4-5

Running from source

git clone https://github.com/claw-sharp/ClawSharp.git
cd ClawSharp

export ANTHROPIC_API_KEY="sk-ant-..."
dotnet run --project src/ClawSharp.Cli

When the REPL starts, type any message and press Enter. Claude will respond in the terminal.

Quick Start: Codex (auth.json)

ClawSharp supports the Codex provider through its native auth.json token file, designed to mirror the original Claude Code behavior.

Locate or create your auth.json

The Codex auth file is typically written by the official codex CLI (~/.codex/auth.json on macOS/Linux, %USERPROFILE%\.codex\auth.json on Windows). It contains an access_token and an account_id.

Example auth.json structure:

{
  "access_token": "eyJ...",
  "account_id": "org-..."
}

ClawSharp reads this file automatically when you use --provider codex or set the Codex environment variables — no extra configuration needed.

Launch with Codex

# Auto-discovers ~/.codex/auth.json
clawsharp --provider codex --model codexplan

# Windows (PowerShell)
clawsharp --provider codex --model codexplan

Using environment variables instead

If you prefer to pass credentials explicitly rather than relying on the auth file:

export CLAUDE_CODE_USE_OPENAI=1
export OPENAI_MODEL=codexplan
export CODEX_API_KEY="eyJ..."         # access_token value
export CODEX_ACCOUNT_ID="org-..."     # account_id value

dotnet run --project src/ClawSharp.Cli

Custom auth file path

export CODEX_AUTH_PATH="/path/to/your/auth.json"
clawsharp --provider codex

Other Providers

| Provider | Environment variable | Example | | ------------- | -------------------- | ------------------------------------------------------ | | Anthropic | ANTHROPIC_API_KEY | clawsharp --provider anthropic | | Gemini | GEMINI_API_KEY | clawsharp --provider gemini --model gemini-2.0-flash | | OpenAI | OPENAI_API_KEY | clawsharp --provider openai --model gpt-4o | | GitHub Models | GITHUB_TOKEN | clawsharp --provider github | | Ollama | (none required) | clawsharp --provider ollama --model llama3.2 |

Full provider setup details: docs/authentication.md

Platform Support

Self-contained binaries are produced for all supported platforms:

| Platform | RID | | ------------------- | ------------- | | Windows 64-bit | win-x64 | | Windows ARM64 | win-arm64 | | Linux 64-bit | linux-x64 | | Linux ARM64 | linux-arm64 | | macOS Intel | osx-x64 | | macOS Apple Silicon | osx-arm64 |

Release archives: .zip (Windows), .tar.gz (Linux/macOS), plus a shared SHA256SUMS checksum file.

Verification

Release archives ship with SHA-256 checksums and GitHub artifact attestations. See docs/distribution.md for verification steps.

Developer Quick Start

git clone https://github.com/claw-sharp/ClawSharp.git
cd ClawSharp

dotnet restore ClawSharp.sln
dotnet build ClawSharp.sln -c Debug
dotnet run --project src/ClawSharp.Cli -- --version

Run all tests:

dotnet test ClawSharp.sln

See docs/contributor-setup.md for a full contributor guide.

Contributing

ClawSharp started as a personal learning project and is still evolving. If you're interested in code agents, terminal tooling, or building more AI-native tools in .NET, contributions, issues, and feedback are welcome.

Documentation

| Document | Description | | ---------------------------------------------------------------------- | ----------------------------------------- | | docs/architect.md | Architecture overview and layer model | | docs/authentication.md | All provider setup and credential options | | docs/user-chat-flow.md | Turn-by-turn conversation flow | | docs/contributor-setup.md | Local development setup | | docs/extension-migration.md | Migrating extensions and plugins | | docs/release-process.md | Maintainer release process | | docs/distribution.md | Distribution, install, and verification | | docs/telemetry-and-diagnostics.md | Debugging and diagnostics | | docs/runbooks.md | Operator support runbooks | | docs/known-gaps.md | Known gaps and deferred work |