Deepx-code is a Go-based, single-binary terminal programming agent designed to enhance developer productivity and reduce LLM token costs. It offers native support for various LLMs (including DeepSeek, MiMo, Kimi, Qwen, and any OpenAI-compatible models) with intelligent routing between cost-effective "flash" and powerful "pro" models. The agent targets developers seeking an efficient, locally-run coding assistant with advanced features like code graph analysis and offline OCR, aiming to significantly cut operational expenses for long-running tasks.

How It Works

Deepx-code employs a unique, zero-token local routing mechanism based on keywords and message length to select between "flash" and "pro" LLM models, minimizing unnecessary token consumption. It features a robust, lossless session persistence using Go's gob binary format, ensuring seamless restarts and state recovery. For long conversations, it implements hierarchical session compression, summarizing older content to stay within context window limits while maintaining high cache hit rates (~99%). Core functionalities like symbol-level code navigation (CodeGraph) and offline OCR (PaddleOCR) are integrated directly, reducing reliance on external services and enhancing local processing capabilities.

Quick Start & Requirements

• Installation: macOS/Linux: curl -fsSL https://raw.githubusercontent.com/itmisx/deepx-code/main/scripts/install.sh | bash && exec $SHELL. Windows (PowerShell): irm https://raw.githubusercontent.com/itmisx/deepx-code/main/scripts/install.ps1 | iex. Gitee mirror available for Chinese users. Installation places deepx in ~/.local/bin/.
• Prerequisites: LLM API keys for chosen providers.
• Dependencies: None explicitly listed beyond standard OS capabilities for the single binary. Sandboxing can optionally use Docker.
• Links: Official Quick Start, Demo GIF.

Highlighted Details

• Single Go Binary: Distributes as a single executable, eliminating Node/Python runtime dependencies for easy deployment across macOS, Linux, and Windows.
• Cost-Effective Caching: Achieves ~99% prompt-cache hit rate for long sessions, drastically reducing costs by avoiding re-processing of repeated context.
• Integrated CodeGraph: Provides symbol-level code navigation, definition lookup, caller/callee analysis, and impact analysis, particularly precise for Go via go/types.
• Local OCR: Includes offline PaddleOCR for recognizing text from screenshots or images without external API calls.
• Workflow Scripting: Supports reusable workflows defined in JavaScript (using agent(), parallel(), pipeline()) for complex, multi-agent task orchestration, compatible with Claude Code's conventions.
• OS-Level Sandboxing: Offers native OS-level isolation (macOS Seatbelt, Linux bubblewrap) or Docker for secure agent execution, with configurable write permissions.

Maintenance & Community

The project is actively maintained by `itmisx`. While specific community channels like Discord or Slack are not prominently featured in the README, the project includes links to multiple language READMEs and a star count, indicating community interest. No major sponsorships or partnerships are listed.

Licensing & Compatibility

The project is licensed under the MIT License, which permits commercial use and integration into closed-source projects without copyleft restrictions.

Limitations & Caveats

The CodeGraph's precision varies by language, with Go receiving the most detailed analysis. While supporting numerous languages, advanced features like precise symbol analysis might be limited for non-Go languages. The "flash" and "pro" model routing is configurable but relies on user-provided API keys and model configurations. The project does not explicitly detail known bugs or alpha status, but offers fallback mechanisms for session loading and sandboxing.