code-assistant

Code Assistant

A CLI tool built in Rust for assisting with code-related tasks.

Features

Autonomous Exploration: The agent can intelligently explore codebases and build up working memory of the project structure.
Reading/Writing Files: The agent can read file contents and make changes to files as needed.
Working Memory Management: Efficient handling of file contents with the ability to load and unload files from memory.
File Summarization: Capability to create and store file summaries for quick reference and better understanding of the codebase.
Interactive Communication: Ability to ask users questions and get responses for better decision-making.
MCP Server Mode: Can run as a Model Context Protocol server, providing tools and resources to LLMs running in an MCP client.

Installation

Ensure you have Rust installed on your system. Then:

1# Clone the repository
2git clone https://github.com/stippi/code-assistant
3
4# Navigate to the project directory
5cd code-assistant
6
7# Build the project
8cargo build --release
9
10# The binary will be available in target/release/code-assistant

Configuration in Claude Desktop

The code-assistant implements the Model Context Protocol by Anthropic. This means it can be added as a plugin to MCP client applications such as Claude Desktop.

Configure Your Projects

Create a file .code-assistant/projects.json in your home directory. This file adds available projects in MCP server mode (list_projects and open_project tools). It has the following structure:

1{
2  "code-assistant": {
3    "path": "/Users/<username>/workspace/code-assistant"
4  },
5  "asteroids": {
6    "path": "/Users/<username>/workspace/asteroids"
7  },
8  "zed": {
9    "path": "Users/<username>/workspace/zed"
10  }
11}

Notes:

The absolute paths are not provided by the tool, to avoid leaking such information to LLM cloud providers.
This file can be edited without restarting Claude Desktop, respectively the MCP server.

Configure MCP Servers

Open the Claude Desktop application settings (Claude -> Settings)
Switch to the Developer tab.
Click the Edit Config button.

A Finder window opens highlighting the file claude_desktop_config.json. Open that file in your favorite text editor.

An example configuration is given below:

1{
2  "mcpServers": {
3    "code-assistant": {
4      "command": "/Users/<username>/workspace/code-assistant/target/release/code-assistant",
5      "args": [
6        "server"
7      ]
8    }
9  }
10}

Usage

Code Assistant can run in two modes:

Agent Mode (Default)

1code-assistant --task <TASK> [OPTIONS]

Available options:

--path <PATH>: Path to the code directory to analyze (default: current directory)
-t, --task <TASK>: Task to perform on the codebase (required unless --continue-task or --ui is used)
--ui: Start with GUI interface
--continue-task: Continue from previous state
-v, --verbose: Enable verbose logging
-p, --provider <PROVIDER>: LLM provider to use [ai-core, anthropic, open-ai, ollama, vertex] (default: anthropic)
-m, --model <MODEL>: Model name to use (defaults: anthropic="claude-3-7-sonnet-20250219", open-ai="gpt-4o", vertex="gemini-1.5-pro-latest", ollama=required)
--base-url <URL>: API base URL for the LLM provider
--tools-type <TOOLS_TYPE>: Type of tool declaration [native, xml] (default: xml) native = tools via LLM provider API, xml = custom system message
--num-ctx <NUM>: Context window size in tokens (default: 8192, only relevant for Ollama)
--agent-mode <MODE>: Agent mode to use [working_memory, message_history] (default: message_history)
--record <PATH>: Record API responses to a file for testing (currently supported for Anthropic and AI Core providers)
--playback <PATH>: Play back a recorded session from a file
--fast-playback: Fast playback mode - ignore chunk timing when playing recordings

Environment variables:

ANTHROPIC_API_KEY: Required when using the Anthropic provider
OPENAI_API_KEY: Required when using the OpenAI provider
GOOGLE_API_KEY: Required when using the Vertex provider
Note: AI Core authentication is configured via deployment config file

Examples:

1# Analyze code in current directory using Anthropic's Claude
2code-assistant --task "Explain the purpose of this codebase"
3
4# Use OpenAI to analyze a specific directory with verbose logging
5code-assistant -p open-ai --path ./my-project -t "List all API endpoints" -v
6
7# Use Google's Vertex AI with a specific model
8code-assistant -p vertex --model gemini-1.5-flash -t "Analyze code complexity"
9
10# Use Ollama with a specific model (model is required for Ollama)
11code-assistant -p ollama -m codellama --task "Find all TODO comments in the codebase"
12
13# Use AI Core provider
14code-assistant -p ai-core --task "Document the public API"
15
16# Use with working memory agent mode instead of message history mode
17code-assistant --task "Find performance bottlenecks" --agent-mode working_memory
18
19# Continue a previously interrupted task
20code-assistant --continue-task
21
22# Start with GUI interface
23code-assistant --ui
24
25# Record a session for later playback
26code-assistant --task "Optimize database queries" --record ./recordings/db-optimization.json
27
28# Play back a recorded session with fast-forward (no timing delays)
29code-assistant --playback ./recordings/db-optimization.json --fast-playback

Server Mode

Runs as a Model Context Protocol server:

1code-assistant server [OPTIONS]