Claude Code Design — Ollama Fork of wieslawsoltes/ClaudeCodeDesign

A space to think. A place to build.

This is a fork of Claude Code Design adapted for use with Ollama — run local models through the same workspace interface without an Anthropic API key.

Built with plain HTML, CSS, JavaScript modules, and a real WebGPU particle renderer. No frontend framework, CDN, or runtime package dependency.

This is not the official Claude Code product. It is not affiliated with or endorsed by Anthropic.

What's changed in this fork

• Ollama support — connect to a local Ollama instance via the built-in proxy (/api/ollama/models, /api/ollama/chat)
• No API key required — select "Local Ollama" in the connect dialog and start chatting
• CORS proxy — scripts/serve.mjs proxies Ollama requests, eliminating browser CORS issues
• NDJSON streaming — handles Ollama's streaming format alongside Anthropic SSE
• Thinking mode toggle — the Think button in the composer turns Ollama reasoning on/off per request (think: true|false); the reasoning trace streams into a collapsible block above the answer

Run locally

Node.js 20 or newer is required. There are no npm dependencies to install.

npm start
# Open http://localhost:4173

Make sure Ollama is running:

ollama serve
ollama pull llama3.2

Connect to Ollama

1. Open Connect API
2. Select Local Ollama
3. Click Connect & test
4. Choose a model in Model settings

The app proxies requests through localhost:4173/api/ollama/* to localhost:11434, so no CORS configuration is needed on the Ollama side.

Connect an Anthropic API key (optional)

The original Anthropic connection is still supported. Open Connect API, enter your key, and choose Connect & test. Keys stay in page memory and are forgotten on reload.

A custom Anthropic-compatible proxy origin is also supported. It must implement GET /v1/models, POST /v1/messages, Anthropic SSE/tool semantics, and CORS.

The experience

Workspace	Working capabilities
Conversations	Live streaming responses, stop/cancel, selected-file context, model discovery, response limits, Build/Plan/Review/Explain modes
Coding partner	Structured list/read/search/propose tools; selected-file access only; a visible activity log
Changes	Per-file proposals, line diffs, explicit approve/reject, stale-base conflict checks, guarded undo
Editor	File creation, editing, renaming, deletion, syntax highlighting, line numbers, local persistence
Preview	Interactive plain HTML/CSS/JavaScript in a sandboxed iframe
Projects	Local files/folders, read-only public GitHub import, source ZIP export, JSON backup/restore
Organization	Searchable sessions, pin/archive/rename/delete, built-in prompt library, command palette

Tests

npm run check
npm test
npm run build
python -m pip install playwright==1.57.0
python -m playwright install chromium

Browser acceptance needs a running server in a separate terminal (the test itself does not start it, and the offline check requires the dist/ build above):

npm start
# Open http://localhost:4173 in another terminal:
python tests/browser.py --url http://127.0.0.1:4173/ --output qa

Deliberate boundaries

This is a browser coding workspace, not full Claude Code CLI parity. It has no shell, npm execution, local OS agent, private GitHub authentication, git writes, background cloud agents, multiplayer collaboration, or deployment of generated projects.

Ollama models support basic chat but not Anthropic-style tool use. The app sends system prompts and file context as plain text.

License

MIT for the original implementation. Product names belong to their respective owners.