Cuckoo Code is a zero-token-cost AI Agent desktop app that embeds AI web UIs in Electron and runs AI-generated tool calls in a local sandbox after confirmation.
- v0.8.22
- Published at
- 5d ago
- 302
- Stars
- 3,145
- Download
- Synced at
- 5d ago
Latest release
v0.8.22Nov 8, 2026, 02:02 AMREADME
Project Overview
Cuckoo Code is a zero-token-cost AI Agent desktop app. Using Electron, it embeds AI web UIs (DeepSeek, Claude, ChatGPT, and more) in a local window and injects a side overlay; a system prompt guides the AI to generate tool call code blocks, which run in a local sandbox after your confirmation and whose results are sent back to the AI. The whole process needs no API key and incurs no per-token API charges, since it uses your web account. It is more than chat: the AI can read and write files, search code, run commands, query databases, and call MCP tools, continuing based on results to form a full Agent loop.
Turn a web AI into a local Agent that can read and write your project and run commands, at zero API cost.
Key Features
- Zero token cost: calls no AI platform API and reuses the web chat capability directly
- Multiple providers: built-in DeepSeek, Claude, and ChatGPT, plus importable custom providers
- A real Agent loop: the AI reads and writes files, searches code, runs commands, queries databases, and calls MCP tools, continuing based on results
- Multi-window management: each window has an independent profile context and can be set to open on launch
- Project initialization: after choosing a project directory, the AI gets the directory tree and system prompt, working against real project context
- Skill support: project-level and user-level skill directories whose name, description, and path are injected at initialization, with full text read on demand
- Agent support: the main conversation can delegate to subagents with isolated context that return only summaries
- Rules support: rules can bind file paths and are injected only when the AI reads matching files
- Tool call system: supports file reading/writing, search, replacement, command execution, web fetching, and database queries
- MCP support: uses the Claude Desktop-compatible format with stdio and http server types
- Overlay panel: shows command previews, execution results, and history, switchable quickly
- Session and compaction: long sessions compact automatically, and login state and settings persist
- Auto retry: retries with backoff when replies are truncated and nudges silent streams with a watchdog
- Safety mechanisms: command and sandbox timeouts, an output buffer limit, and confirmation for dangerous commands
Target Audience
- Developers who want to drive local projects with a web AI
- Users who want to avoid API billing
- Advanced users who want to extend AI capabilities with Skills, Agents, and Rules
Quick Start
- Download the latest installer from Releases, or run it from source
- Launch the app and choose a platform (DeepSeek, Claude, or a custom provider)
- Log in to the corresponding platform's web account as usual
- Click "Initialize Project" and choose a project directory; the AI receives the directory tree and system prompt
- Chat with the AI and ask it to modify files, run commands, or query code
- Tool calls in the AI's replies are detected automatically, run in the sandbox after confirmation, and their results are sent back to continue
Core Advantages
- No API cost: reuses your web account, bound only by the web service's own limits with no per-token interface charges
- Actually does the work: upgrades from chat to an Agent that reads and writes files and runs commands for real development tasks
- Aligned mechanisms: Skill, Agent, and Rules follow Claude Code's directory conventions for low migration cost
- Controllable context: subagents isolate context and long sessions compact automatically, easing slowdowns in very long chats
System Requirements
| Item | Requirement |
|---|---|
| OS | Windows, macOS |
| Runtime | Node.js 16.0.0+ (18+ recommended) and npm when running from source |
| Account | A working web account for the AI platform you use |
Notes
- The app ships no AI of its own; you must log in to the corresponding platform's web account before use
- Tools execute inside your local project directory, so review each tool call carefully
- Dangerous commands require confirmation, but do not let the AI run destructive operations in important directories
- If electron and esbuild postinstall scripts are blocked during dependency install, approve them and reinstall
- This is a third-party tool and is not affiliated with any AI platform it connects to