Code Assistant MCP server
Commit Nov 13, 2026Lets an AI client run an autonomous coding agent that reads, searches, edits files and runs commands in your project.
Also known as stippi/code-assistant
- Stars
- 190
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 13, 2026
- License
- MIT
What is Code Assistant MCP?
Code Assistant lets an AI client run an autonomous coding agent that reads, searches, edits files and runs commands in your project.
Summary generated from the project's README
About Code Assistant MCP server
Running an autonomous coding agent that manages your codebase directly from an AI client is what this tool enables. It functions as a bridge that allows models to read, search, edit files, and execute commands within your local project. Developers can use it through a native GUI, a terminal interface, or as a headless MCP server for clients like Claude Desktop. It also supports an ACP agent mode for editors such as Zed.
To get started, you select an LLM provider like Anthropic, OpenAI, or Google Vertex AI and provide your API key through a settings screen. The system is designed to be transparent, showing exactly which tools are running and what context they return. It handles complex tasks like reconciling file edits with auto-formatters and reading non-code documents like PDFs or Excel files by converting them to Markdown. This ensures the agent maintains an accurate mental picture of your project without wasting tokens on unnecessary re-reads.
What you can do with it
- 1Automated codebase explorationAsk the agent to explain the structure and logic of a complex repository.
- 2Document-integrated developmentDirect the agent to use project documentation like PDFs or Word files to inform its coding decisions.
- 3Web application testingHave the agent drive a real browser to navigate and interact with your web application.
- 4Context-aware code editingPerform file edits that remain consistent even when your project uses automatic code formatting.
Written from the project's README. Every use case is checked against a quote from it.
Configuration
ANTHROPIC_API_KEY
| Name | Description | |
|---|---|---|
| ANTHROPIC_API_KEY | — |
Automated checks
Ran Nov 12, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablePass
- Open source license (OSI)Pass
- Commit or release in the last yearPass
- Package name not close to a popular packageNot checked
- Remote endpoint uses HTTPSNot checked
- Destructive tools are labelledNot checked
- No hidden instructions to the modelPass
- No critical advisory in the package or its dependenciesNot checked
README
View on GitHubCode Assistant
An open-source AI coding agent, written in Rust, with a native GUI, a terminal mode, and integrations for editors and MCP clients. It runs an autonomous agent loop over your codebase — reading, searching, editing files, and running commands — while keeping you in the loop about what it is actually doing.
Getting started
The quickest way to try it is a prebuilt download — no toolchain required:
- Grab the latest build from the Releases page:
- macOS:
Code-Assistant-macos-aarch64.app.zip(Apple Silicon) orCode-Assistant-macos-x86_64.app.zip(Intel) - Linux:
code-assistant-linux-x86_64.zip - Windows:
code-assistant-windows-x86_64.zip
- macOS:
- Launch it. On first start, code-assistant opens its Settings screen. Pick a provider (there are one-click suggestions for the common ones), paste an API key, and choose a model — that's it.
- Add your project folder from within the app and start a chat.
No hand-editing of JSON files required to get going. (You still can, if you prefer — see the Configuration guide.)
# Install the Rust toolchain (macOS/Linux)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
# On Linux (Debian/Ubuntu), install the system libraries gpui needs:
sudo apt-get install -y --no-install-recommends \
pkg-config build-essential libssl-dev libzstd-dev \
libfontconfig-dev libwayland-dev libx11-xcb-dev \
libxkbcommon-x11-dev libasound2-dev libvulkan1
# On macOS, install the metal toolchain:
xcodebuild -downloadComponent MetalToolchain
# Clone and build
git clone https://github.com/stippi/code-assistant
cd code-assistant
cargo build --release
The binary lands at target/release/code-assistant. See the
Configuration guide for building a self-contained
macOS .app bundle.
What makes it different
A handful of things we care about:
- A UI you can actually follow. Instead of a terse "Called 5 tools", code-assistant shows which tools ran and what each of them handed back to the model as context. When it goes off the rails, you can see why.
- Format-on-save that stays token-efficient. When your project auto-formats code, an agent's mental picture of a file goes stale and its follow-up edits start failing for non-obvious reasons. code-assistant runs your formatter and then reconciles the model's own edits with the formatted result — without wasting tokens on re-reading files. See docs/format-on-save-feature.md.
- Transparent file encoding & line endings. It reads a file however it is stored (encoding, BOM, CRLF/LF), gives the model clean text, and writes it back the way it was — so it never silently rewrites your line endings.
- Searches and reads documents, not just code. Word, Excel, PowerPoint, PDF and more are consumed automatically as Markdown, so you can point the agent at real-world documents alongside your source.
Features
- Multiple LLM providers: Anthropic, OpenAI, Google Vertex AI, Ollama, OpenRouter, SAP AI Core, Groq, Cerebras, Mistral, and more.
- Four ways to work: a native GUI (built on Zed's GPUI), a terminal mode, a headless MCP server, and an ACP agent for editors like Zed.
- Adaptive tool syntax: native function calling, XML tags, or triple-caret blocks — chosen per session to fit the model.
- Real-time streaming with smart filtering that blocks unsafe tool combinations (e.g. editing a file before reading it).
- Sessions per project with branching, persistent state, and draft messages with attachments.
- Sub-agents, permission tiers, a command sandbox, and automatic context compaction when the window fills up.
- MCP client mode: plug in external Model Context Protocol servers and use their tools.
- Browser sessions: the agent drives a real browser (navigate, read, click, type) to test web apps, with a human-in-the-loop login handoff so it acts as you without ever seeing your credentials.
- Skills: reusable, task-specific playbooks the agent can load on demand.
- Auto-loaded guidance: picks up
AGENTS.md(orCLAUDE.md) from your project root to align with repo-specific instructions.
Interfaces
code-assistant # native GUI (default)
code-assistant --tui # terminal interface
code-assistant acp # ACP agent for editors like Zed
code-assistant server # headless MCP server (e.g. Claude Desktop)
Any mode can take an initial task: code-assistant --task "Explain this codebase".
Add to your Zed settings:
{
"agent_servers": {
"Code-Assistant": {
"command": "/path/to/code-assistant",
"args": ["acp", "--model", "Claude Sonnet 4.5"],
"env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
}
}
}
See Zed's docs on custom agents.
In Claude Desktop settings (Developer tab → Edit Config):
{
"mcpServers": {
"code-assistant": {
"command": "/path/to/code-assistant",
"args": ["server"],
"env": { "SHELL": "/bin/zsh" }
}
}
}
Configuration
For most people the in-app Settings screen is enough — it manages providers, models, MCP servers and skills for you. Everything can also be configured via JSON files, and there are more advanced options (project setup, sandbox modes, tool syntax, session recording, CLI flags):
→ See the Configuration guide.
Contributing
Contributions are welcome! The codebase is a decent tour of async Rust, AI agent architecture, and cross-platform UI. If something about the agent's behaviour annoys you, that is exactly the kind of detail this project cares about: please open an issue.
Roadmap
Not really a roadmap — just a few directions, in no particular order:
- Block replacing in stale files: detect early when the model is about to
replace_in_filea file that changed since it last read it, and reject with a helpful message. - Compact tool-use failures: strip failed tool calls / mismatched search blocks from the history to save tokens.
- Memory tools: help the agent build up a knowledge base for a project.
- Tighter sandboxing: restrict tool access to git-tracked files within the project.
- Fuzzy matching of search blocks: reduce the re-reads caused by failed exact matches.
Code Assistant MCP questions
What can the Code Assistant MCP server do?
It runs a headless MCP server that gives an AI client an autonomous coding agent loop over your codebase, reading, searching, editing files and running commands.
How do I connect Code Assistant to Claude Desktop?
Add an mcpServers entry pointing the command at the code-assistant binary with the 'server' argument, optionally setting the SHELL environment variable.
Which LLM providers does Code Assistant support?
It supports Anthropic, OpenAI, Google Vertex AI, Ollama, OpenRouter, SAP AI Core, Groq, Cerebras, Mistral and others.
Can Code Assistant use external MCP servers?
Yes, it has an MCP client mode so you can plug in external Model Context Protocol servers and use their tools.
Does Code Assistant work with editors like Zed?
Yes, it provides an ACP agent that you register in Zed's agent_servers settings.