Loading repository data…
qinghaoliqaq / repository
An AI-powered desktop coding workspace for planning, coding, debugging, review, and testing through multi-agent orchestration and shared blackboards.
AutoCoder is an AI-powered desktop coding workspace for end-to-end software delivery. It combines a Tauri desktop shell, a React frontend, and a Rust orchestration backend to coordinate planning, coding, debugging, review, and testing across multiple agents.
This repository is currently an early-stage open-source project. The internal product name AI Dev Hub still appears in parts of the UI and codebase.
Most AI coding tools are optimized for one agent in one chat. AutoCoder is built around a different idea: software delivery is a workflow, not a single prompt.
This project tries to make that workflow explicit:
Director decides which skill should run nextplan, code, debug, review, and test are treated as separate execution modesPLAN_BLACKBOARD.md, BLACKBOARD.md, BLACKBOARD.json) act as the coordination layerIn practice, AutoCoder is trying to be a local AI software delivery console rather than a plain chat wrapper.
The Director is the traffic controller. It reads the user's input and decides whether the app should:
The Director does not implement features itself. It routes work to the right skill.
The backend organizes execution into separate skills:
plan: creates or reviews a technical plancode: implements subtasks against PLAN.md using a shared blackboard loopdebug: focuses on fault isolation and repairreview: checks implementation coverage, security, and cleanuptest: generates test plans, runs integration checks, and writes project reportsInstead of letting agents coordinate through hidden conversation alone, AutoCoder writes structured coordination state into project files. That makes the workflow more inspectable and easier to recover.
Typical files:
PLAN.mdPLAN_BLACKBOARD.mdBLACKBOARD.mdBLACKBOARD.jsonchange.logbugs.mdThe Rust backend shells out to local CLIs for execution:
claudecodexThese runners stream tokens back into the UI, record tool activity, and support cancellation.
AutoCoder is split into three layers:
Desktop shell: Tauri windowing, native capabilities, and app lifecycleFrontend UI: chat, history, tool logs, blackboard views, file explorer, and status panelsRust orchestrator: Director routing, workspace access, session persistence, prompt loading, skill execution, and CLI runnersHigh-level flow:
User
-> Director
-> plan / code / debug / review / test
-> Claude / Codex runners
-> blackboard files + session history + tool logs
-> desktop UI
The important design choice is that the workflow state is externalized into files and history, rather than hidden inside one long agent conversation.
.
├── src/ # React UI
├── src-tauri/ # Rust backend and Tauri app
│ ├── src/
│ │ ├── director.rs # Director orchestration
│ │ ├── history.rs # Session persistence
│ │ ├── workspace.rs # Workspace and file access
│ │ └── skills/ # plan/code/debug/review/test runners
│ └── prompts/ # Prompt templates used by the backend
├── config.example.toml # Example Director configuration
Before running the app locally, you need:
claude CLI installed and available in PATHcodex CLI installed and available in PATHconfig.toml or environment variablesCopy the example config:
cp config.example.toml config.toml
Then fill in your Director settings:
[director]
api_key = "your-api-key"
base_url = "https://api.openai.com/v1"
model = "gpt-4o"
api_format = "openai"
Supported API wire formats:
openaianthropicThe app also supports environment variable overrides:
DIRECTOR_API_KEYDIRECTOR_BASE_URLDIRECTOR_MODELDIRECTOR_API_FORMATInstall JavaScript dependencies:
npm install
Start the frontend dev server:
npm run dev
Start the desktop app:
npm run tauri dev
Build the frontend bundle:
npm run build
Run frontend unit tests:
npm test
Run Rust tests:
cd src-tauri
cargo test
After configuration, a typical local flow is:
npm run tauri devplan generate PLAN.mdcode execute subtasks through the blackboard loopreview and test to verify the resultIf the local claude or codex CLIs are missing, the app will detect that and show the missing-tool state in the UI.
AutoCoder is usable as a serious prototype, but it is still in active development.
Current characteristics:
AI Dev Hub identityIf you are evaluating this project, treat it as an open-source alpha rather than a finished product.
AI Dev HubThis repository includes a vendored skill architecture for selectively reusing external skill assets without relying on global agent skill discovery.
Key rule:
See VENDORED_SKILLS_ARCHITECTURE.md for the design rationale.
Contributions are welcome, especially in these areas:
If you open an issue or PR, include:
plan, code, debug, review, or testThis project is licensed under the Apache License 2.0. See LICENSE.
Selected from shared topics, language and repository description—not editorial ratings.
nevoflux-browser /
AI-powered browser & computer automation agent in Rust — drives a real browser and the desktop, speaks MCP and an OpenAI-compatible API, runs WASM skills, and ships as a headless Docker service.
BMOBOX /
AI-powered code editor using an **agent swarm** where multiple AI agents collaborate to generate, refactor, debug, and understand code. Built with **Tauri, Rust, and React** for a fast native desktop experience.