DreamCoder (GoDiao/dreamcoder)
DreamCoder (GoDiao/dreamcoder) is an open-source, locally run AI coding desktop workspace built on Electron, React 19, and Bun. It combines Claude Code workflows, visual MCP server configuration, a built-in PTY terminal (xterm.js), an integrated side-panel browser, a persistent cross-worktree Project Board, /design artifact prototyping (60+ templates, 156+ design systems), and optional LAN H5 Access to continue desktop sessions from a phone. Presets cover Anthropic and OpenAI-compatible endpoints including DeepSeek, Qwen, Kimi, Zhipu GLM, LM Studio, and Ollama.
1. Core workspace capabilities & MCP / Computer Use stack
DreamCoder replaces terminal-only coding loops with a tabbed local workspace while preserving transparent tool calls and git worktree hygiene:
| Capability | What DreamCoder Provides | Documented Notes & Scope |
|---|---|---|
| Multi-provider routing | Anthropic + OpenAI-compatible endpoints with visual connection/latency checks | Built-in presets for DeepSeek, Qwen, Kimi, Zhipu GLM, LM Studio, Ollama, plus custom endpoints |
| Visual MCP extensions | Manage MCP servers in Settings UI without hand-editing JSON files | Supports WebSocket, MCP, and LSP protocols alongside Claude Code workflows |
| Built-in browser | Side-panel tab for local dev servers, project HTML, or HTTPS sites (persistent sign-ins) | Desktop app only; agent asks before acting on external sites unless trusted under Settings -> Permissions |
| Computer Use | Background desktop app and browser UI automation powered by cua-driver (MIT) | Auto-downloaded when enabled (no Python required); verified on macOS, community-tested on Windows, preview on Linux |
Artifacts & /design | Persistent dc_artifacts/ pages, canvases, website cloning, and interactive ShowAI pages | 60+ templates and 156+ design systems (converted from Open Design, Apache-2.0); PDF/PNG/HTML export |
| Project Board & LAN H5 | Shared task/sub-task board across sessions/worktrees with owner & reviewer gates; QR-code phone access | H5 Access requires desktop app running on same LAN; cross-network access requires your own reverse proxy |
2. Install from Releases or build the Bun monorepo from source
Download an unsigned pre-built installer for macOS (Apple Silicon) or Windows (x64) from GitHub Releases, or run the development build from source with Bun >= 1.0 and Node.js >= 18:
# 0. Clone the repository
git clone https://github.com/GoDiao/dreamcoder.git
cd dreamcoder
# 1. Install root sidecar dependencies and ShowAI page reader
bun install
(cd vendor/showai && bun install)
# 2. Install desktop dependencies (Electron + React 19 frontend)
cd desktop && bun install
# 3. Download the Electron binary (Bun skips Electron's install script)
bun run electron:install
# 4. Build the sidecar binary
bun run build:sidecars
# 5. Launch desktop app in dev mode (Vite + Electron)
bun run electron:dev
To package a local installer, run bun run electron:pack (outputs to desktop/electron-release/). All five source steps are required; electron:dev fails if either the Electron binary or sidecar binary has not been built first.
3. Platform support & architecture overview
| Platform | Maintainer Status | Pre-built Installer |
|---|---|---|
| macOS arm64 | Maintainer's daily platform (.dmg since v0.5.0) | Yes (unsigned .dmg) |
| Windows x64 | Extensively tested former primary platform | Yes (unsigned -setup.exe) |
| Linux x64 | Best-effort from source; not in daily CI validation | No pre-built installer |
Under the hood, DreamCoder pairs an Electron shell and React 19 + Vite + TailwindCSS 4 + Zustand frontend with a Bun sidecar runtime (PTY terminal via xterm.js, MCP/LSP/WebSocket protocols, and ShowAI interactive page rendering).
4. Security boundaries & related Agent Skills Hub guides
- Unsigned binaries & permission boundaries: Review release notes before bypassing OS Gatekeeper/SmartScreen prompts on unsigned builds, and keep external browser domains untrusted until needed under
Settings -> Permissions. - Local storage vs. cloud inference: API keys and settings remain in local files, while prompts sent to remote providers follow those providers' policies (use Ollama or LM Studio presets for fully local inference).
- Next steps: Audit any server you add in DreamCoder's MCP panel with our MCP server security checklist, browse vetted servers in Best MCP servers, or compare multi-provider workspaces in Claude Code vs Codex vs Cursor.
Frequently asked questions
Does DreamCoder support local open-weight models like Ollama or LM Studio?
Yes. DreamCoder supports Anthropic and OpenAI-compatible endpoints with built-in provider presets for Ollama, LM Studio, DeepSeek, Qwen, Kimi, and Zhipu GLM, plus custom endpoints with an in-app latency and connectivity test.
How does DreamCoder manage MCP servers?
DreamCoder includes a visual MCP Extensions panel in Settings so you can add, configure, and toggle MCP servers in the desktop UI instead of hand-editing raw JSON config files.
Does Computer Use in DreamCoder require installing Python?
No. When you enable Computer Use in Settings, DreamCoder automatically downloads cua-driver (MIT) without requiring a local Python environment. Computer Use is currently verified on macOS, needs community testing on Windows, and is preview-stage on Linux.
Can I access DreamCoder sessions from my phone outside my home network?
Built-in H5 Access works over the same LAN while the desktop app stays running (using a token and QR code). Reaching sessions across different networks requires configuring your own reverse proxy.