gray-code

GrayCode 2.0 is a local-first AI workbench and coding assistant, offering a standalone desktop application (Windows x64) that supports understanding the workspace, searching and modifying code, executing commands, invoking language services, and reviewing file changes via Diff. It supports multi-model conversations, role-based chat, bot entry points, and sub-agent collaboration with shared tasks. Core features include: - **Multi-model channels**: Supports Gemini, OpenAI Chat Completions-compatible interfaces, OpenAI Responses, and Anthropic; each channel can independently configure models, tool modes, thinking, retries, and token counting. - **Real code operations**: Read, search, and modify files, run terminal commands, invoke language services, and support multimodal context such as images and PDFs; writes can be reviewed via Diff. - **Structured workflows**: Built-in Design, Plan, Progress, Review, and TODO tools make complex tasks traceable from proposal to verification. - **Extensible agent capabilities**: Connect to MCP Servers, load reusable Skills, collaborate via foreground or background Sub-Agents, and also connect to external coding agents that support ACP, such as Kimi Code. - **Computer and browser operations**: Browse and operate web pages in the built-in browser, and also let the model operate Windows applications to complete tasks that require a graphical interface. - **Local permanent memory**: Global and workspace memory are isolated from each other, preserving conventions, knowledge, and decisions across sessions without relying on external memory services. - **Long tasks and long conversations**: Supports message queues, automatic summarization, checkpoints, background result回流, as well as token, cost, and usage time statistics. - **Tree-branch conversations**: Retries and edits no longer overwrite old answers; each candidate branch can be switched and continue developing independently, with optional linkage to workspace checkpoints when switching. In addition, GrayCode provides dedicated image preprocessing for DeepSeek vision models, including page-by-page PDF rasterization, GIF animation frame splitting, official format normalization, image size optimization, and pre-request validation. The project provides a Windows desktop version (portable and installer versions), and the 1.x VS Code extension source code is retained in the `v1-extension` branch. All core data is stored locally, and users can freely connect different model channels. Documentation includes a user manual, development environment and contribution instructions, project structure, and version change records. The license is AGPL-3.0-only, with a scope-limited Cubism combined license exception.