cdesktop: Open-Source Claude Code Desktop, Any Provider

If you have been using the official Claude Code CLI and wishing you could point it at a different model, run it fully offline, or just wrap it in a proper desktop window, cdesktop is the answer. This open-source project gives you a native desktop experience for agentic AI coding, launches instantly via npx, and lets you swap in any OpenAI-compatible provider, including local models running on your own hardware.

No waitlist. No locked-in subscription. No telemetry. Just run one command and start coding.

What Is cdesktop, Exactly?

cdesktop is a community-built, open-source desktop shell for the Claude Code agent runtime. Think of it as a thin but polished native window that wraps the same agentic loop Claude Code uses under the hood, while stripping out the provider lock-in and adding provider flexibility.

The core value proposition is threefold:

1. Provider agnostic. The official Claude Code Desktop only talks to Anthropic’s API. cdesktop accepts any endpoint that follows the OpenAI chat completions spec, which now covers virtually every major model provider and local runtime.

2. Zero-install entry point. Because it ships as an npm package, you can start it with a single npx cdesktop command without ever globally installing anything. Node.js is the only prerequisite.

3. Open source. The full codebase is public. You can audit it, fork it, self-host it, or extend it with custom tools and system prompts.

If you have spent time with Claude’s agentic coding capabilities or pushed the limits of what the CLI can do, cdesktop removes the biggest remaining friction point: the requirement to use Anthropic’s hosted API exclusively.

Getting Started: Running cdesktop via npx

The fastest path from zero to a working AI coding agent desktop is three steps.

Prerequisites

You need Node.js 18 or later. Check your version with:

node --version

If you are behind, grab the latest LTS from nodejs.org.

You also need an API key from whichever provider you plan to use. If you are going the fully local route with Ollama, you need Ollama installed and at least one model pulled.

Step 1: Launch cdesktop

npx cdesktop

The first run downloads the package from npm (roughly 20-30 MB) and opens the desktop window. Subsequent runs use the cached version unless a new release has been published.

If you want to pin a specific version for reproducibility:

npx cdesktop@latest

Or install it globally to skip the npx overhead on repeated launches:

npm install -g cdesktop
cdesktop

Step 2: Configure Your Provider

On first launch, cdesktop drops you into a configuration screen. The key fields are:

For Anthropic (the default), you just paste your API key and choose a model. For other providers, you swap out the Base URL.

Step 3: Open a Project

Click “Open Folder” and point cdesktop at any local repository. The agent immediately gets read access to the file tree and can start reading, editing, and running code in that context.

Connecting Any Provider: Practical Config Examples

This is where cdesktop earns its “any provider” claim. Below are the exact config values for the most popular setups.

OpenAI GPT-4o

Base URL: https://api.openai.com/v1
Model:    gpt-4o
API Key:  sk-...

OpenAI’s API follows the same chat completions spec cdesktop targets, so the tool calls and streaming work without any adapter layer.

Local Ollama (Zero Cost, Zero Telemetry)

If you have Ollama running locally with a model like qwen2.5-coder:32b, the config is:

Base URL: http://localhost:11434/v1
Model:    qwen2.5-coder:32b
API Key:  ollama   (any non-empty string works)

No internet connection required after the model is pulled. This is the most privacy-friendly setup and eliminates API costs entirely. For hardware and quantization guidance on running large local models, our Qwen 27B VRAM guide covers the backend tradeoffs in detail.

LM Studio

LM Studio’s local server also exposes an OpenAI-compatible endpoint:

Base URL: http://localhost:1234/v1
Model:    (match the model name shown in LM Studio)
API Key:  lm-studio

Groq (Fast Inference)

Base URL: https://api.groq.com/openai/v1
Model:    llama-3.3-70b-versatile
API Key:  gsk_...

Groq’s inference speeds make it compelling for interactive coding agents where latency matters more than raw capability.

cdesktop vs Official Claude Code Desktop: Feature Comparison

Let’s be specific about what you gain and what you trade away compared to Anthropic’s official product.

The honest summary: if you are primarily using Claude models and want the most polished, supported experience, Anthropic’s official tooling wins. If you want flexibility, zero cost, local execution, or provider experimentation, cdesktop is the more powerful choice.

For a deep comparison of the underlying models themselves, see our Claude vs ChatGPT coding benchmarks.

Advanced Usage: Custom System Prompts and Agent Tuning

One of the underappreciated features of cdesktop is how much control you get over the agent’s behavior through custom system prompts. The official Claude Code has a hardcoded persona. cdesktop lets you override it entirely.

Writing an Effective Coder System Prompt

A minimal but effective custom system prompt for coding agents:

You are an expert software engineer working inside the user's codebase.

Rules:
1. Always read the relevant files before editing them.
2. Make the smallest diff that achieves the goal.
3. Run tests after making changes if a test command is available.
4. When uncertain about intent, ask one clarifying question before proceeding.
5. Never delete code without confirming the user does not need it.

The key principles from advanced prompt engineering techniques apply directly here: be specific about constraints, define the tool loop behavior explicitly, and anchor the agent to the user’s intent rather than its own assumptions.

Adjusting Max Tokens for Cost vs. Depth

Agentic coding loops burn tokens fast because the model reads files, plans, edits, and verifies in a multi-step loop. A reasonable starting point:

• Quick edits / small files: 4096 tokens
• Feature implementation: 8192 tokens
• Large refactors: 16384 tokens or higher (check provider pricing)

With local Ollama models, this constraint disappears entirely since you are not paying per token.

Project-Level Context

cdesktop reads your working directory, but the agent performs better when you provide a AGENTS.md or CLAUDE.md file in the repo root with project-specific context. Include:

• The tech stack and framework versions
• Conventions (naming, testing, commit style)
• Commands to run tests, lint, and build
• Anything the agent should never do (e.g., “never modify the database migration files directly”)

This is the same technique that powers the most effective setups described in our 8 Advanced Claude Code Tips guide.

Real-World Use Cases Where cdesktop Shines

Enterprise and Compliance Environments

Many teams cannot send source code to external APIs. cdesktop with a local Ollama backend keeps every byte on-premise. Pair it with a self-hosted open-weight model like DeepSeek Coder or Qwen2.5-Coder and you have a fully air-gapped AI coding agent.

Cost-Conscious Development

Claude Sonnet’s API pricing is reasonable for occasional use, but agentic loops compound token usage fast. Running a local 32B coder model via Ollama costs exactly $0 per query after the initial hardware investment. For teams running hundreds of agentic tasks per day, this math matters significantly.

Multi-Model Experimentation

If you are evaluating models for a production workflow, cdesktop lets you benchmark the same coding task against Claude, GPT-4o, and a local Qwen model without switching tools. Just update the provider config and re-run. This kind of systematic comparison is exactly how you build intuition about which AI model is right for your use case.

Learning and Prototyping

For developers learning how agentic AI coding works, cdesktop’s open-source codebase is a reading list. You can trace exactly how the agent loop constructs tool calls, handles file edits, and processes the model’s responses. It is a working reference implementation.

Troubleshooting Common Issues

The agent stalls and never takes an action. Most likely cause: the provider does not support tool calling, or the model you selected does not implement it. Switch to a model with confirmed tool call support (GPT-4o, claude-sonnet-4-5, or llama-3.3-70b on Groq).

npx is slow on first launch. The npm registry download takes 15-30 seconds on the first run. Use npm install -g cdesktop if you want instant subsequent launches.

Ollama connection refused. Ensure Ollama is running (ollama serve) before launching cdesktop. Default port is 11434. Check with curl http://localhost:11434/v1/models.

The agent edits the wrong file. Open a tightly scoped folder rather than your entire home directory. The agent uses the working directory as its context window for file discovery. Smaller scope means more precise edits.

API key errors with local providers. Some local providers require a non-empty but otherwise arbitrary API key. Use the string local, ollama, or lm-studio as a placeholder: cdesktop passes it through without validation.

Should You Switch From Official Claude Code?

Not necessarily “switch,” more like “add to your toolkit.” The two are complementary:

Use official Claude Code / Claude Code Desktop when you want maximum reliability, Anthropic’s latest model capabilities (including claude-opus-4 and claude-sonnet-4-5’s full 200K context), and first-party support. For serious production coding workflows tied to Claude specifically, the official tooling is the right choice.

Use cdesktop when you need provider flexibility, want zero-cost local execution, are in a compliance-sensitive environment, or are experimenting with open-weight models for a specific task type.

The best developers in 2026 are not loyal to a single tool. They understand what each tool optimizes for and choose accordingly. For a broader map of where Claude Code fits in the AI coding assistant landscape, our best AI coding assistants roundup covers the full competitive field.

The Bottom Line

cdesktop is one of the cleanest examples of the open-source ecosystem filling a gap the official product deliberately leaves open. Anthropic’s Claude Code is excellent if you are using Anthropic’s API. cdesktop is excellent if you are not, or if you want the freedom to be.

The npx cdesktop entry point removes every barrier. No account, no install, no commitment. Run it once against a local Ollama model and you have a fully functional, zero-cost AI coding agent that never sends your code anywhere. That is a genuinely different product from what any of the major labs are shipping.

For developers building serious agentic workflows, this kind of provider-agnostic tooling is not a curiosity: it is infrastructure. Start with one command. See what your local models can actually do with proper agentic scaffolding. Then decide where you want to invest.

Try it now: Run npx cdesktop in your project directory. The cdesktop GitHub repository has full documentation, provider setup guides, and an active issues tracker for community support.

If you are building your own tools on top of Claude’s API rather than using pre-built clients, check out our step-by-step guide to building your first AI agent with the Claude API.