code-assistant is an open-source, LLM-powered autonomous coding assistant written in Rust, offering a native GUI, a terminal mode, and integrations for editors and MCP clients. It runs an autonomous agent loop on your codebase—reading, searching, editing files, and running commands, while keeping you informed about what it is actually doing at all times.
Key capabilities
- Followable interface: instead of showing a vague "called 5 tools", it shows which tools ran and what context each tool returned to the model, making it easy to pinpoint the cause when something goes wrong.
- Format-on-save and token-efficient: when a project auto-formats code, the agent's knowledge of the file becomes stale, and subsequent edits may fail for non-obvious reasons. code-assistant runs your formatter, then reconciles the model's own edits with the formatting result, without wasting tokens re-reading the file.
- Transparent file encoding and line endings: reads files as they are stored (encoding, BOM, CRLF/LF), provides clean text to the model, then writes back as-is, without silently rewriting line endings.
- Search and read documents, not just code: Word, Excel, PowerPoint, PDF, etc. are automatically consumed as Markdown, so you can have the agent work with real documents and source code at the same time.
- Multiple LLM providers: Anthropic, OpenAI, Google Vertex AI, Ollama, OpenRouter, SAP AI Core, Groq, Cerebras, Mistral, and more.
- Four ways of working: native GUI (based on Zed's GPUI), terminal mode, headless MCP server, and ACP agent for editors such as Zed.
- Adaptive tool syntax: native function calling, XML tags, or triple-backtick blocks—chosen per session to suit the model.
- Real-time streaming, with smart filtering that blocks unsafe tool combinations (for example, editing a file before reading it).
- Per-project sessions, with branching, persistent state, and draft messages with attachments.
- Sub-agents, permission tiers, command sandboxing, and automatic compaction when the context window fills up.
- MCP client mode: connect to external Model Context Protocol servers and use their tools.
- Browser sessions: the agent drives a real browser (navigate, read, click, type) to test Web application and provides human-in-the-loop login handoff, allowing it to act as you without seeing your credentials.
- Skills: reusable, task-specific playbooks that the agent can load on demand.
- Automatic loading of guidance: reads AGENTS.md (or CLAUDE.md) from the project root to align with repository-specific instructions.
How to use
The fastest way to try it is to download a prebuilt version, with no toolchain required:
- macOS: Code-Assistant-macos-aarch64.app.zip (Apple Silicon) or Code-Assistant-macos-x86_64.app.zip (Intel)
- Linux: code-assistant-linux-x86_64.zip
- Windows: code-assistant-windows-x86_64.zip Get the latest build from the Releases page. After launching, the first run opens the settings interface. Choose a provider (common providers have one-click suggestions), paste the API key, and select a model. Add a project folder in the app and start chatting. You can get started without manually editing JSON files (if you prefer, you can, see the configuration guide). When building from source, the binary is located at target/release/code-assistant. For building a self-contained macOS .app bundle, see the configuration guide. Any mode can accept an initial task: code-assistant --task "Explain this codebase".
Editor and client integration
- Zed: add to Zed settings, see Zed's documentation on custom agents.
- Claude Desktop: configure in settings (Developer tab → Edit Config). For most people, the in-app settings interface is enough—it manages providers, models, MCP servers, and skills for you. Everything can also be configured through JSON files, and there are more advanced options (project settings, sandbox mode, tool syntax, session logs, CLI flags), see the configuration guide.
When to use it Suitable for developers who need to run autonomous agent loops on a codebase while wanting a clear understanding of its behavior; suitable for users who use multiple LLM providers and need to work in a GUI, terminal, editor, or MCP client; also suitable for scenarios where the agent needs to handle both code and documents such as Word, Excel, PowerPoint, and PDF, or where browser testing of Web applications is needed.
Limitations The materials do not provide specific details on pricing, platform support scope, performance results, or usage steps. The project welcomes contributions, and the codebase involves asynchronous Rust, AI agent architecture, and cross-platform UI. The project lists some directions (not a roadmap): chunk replacement for stale files, compression tool usage failures, memory tools, stricter sandboxing, fuzzy matching for search chunks.
