DeepCode - HKUDS | ai.dosa.dev

DeepCode

HKUDS · AI-Native IDEs & Editors · Updated Aug 12, 2026

DeepCode is an open-source, autonomous AI coding agent designed for complex repository-level tasks, scientific research reproduction, and automated software engineering. It utilizes a multi-stage framework that includes blueprint distillation, stateful memory (CodeMem), and closed-loop verification to manage information flow effectively across multi-file projects.

Recent Updates

Introduced a V2 architecture with a unified agent runtime (CLI, Desktop, and Headless), persistent sessions, security sandboxing, and improved multi-agent collaboration via MCP.

As of Aug 12, 2026

Key Features

Autonomous agent loop for multi-step engineering tasks
Paper2Code research-to-implementation pipeline
Stateful Code Memory (CodeMem) for cross-file consistency
Automated verification for runtime error correction
Support for custom Skills and lifecycle hooks

Who It's For

✓ Best For

Complex, multi-step software engineering tasks, reproducing research papers from technical documents, and repository-level refactoring.

✗ Not Ideal For

Simple, one-off code completions or real-time autocomplete tasks where low-latency IDE-native suggestions are required.

Pricing

Free

Open-source software; users provide their own API keys for model providers (e.g., DeepSeek, OpenAI, Anthropic, or local models), paying usage-based costs directly to those providers.

Verdict

A powerful, research-backed agent for autonomous, long-horizon coding tasks, though it requires more manual setup and explicit task definition compared to commercial IDE-native assistants.