CodeCompass helps developers tackle legacy or existing codebases by giving AI coding assistants the context they need to deliver spot-on suggestions. Legacy code is tough for AI—it’s often messy, outdated, and lacks clear documentation. CodeCompass solves this by analyzing your codebase with Qdrant Vector Store and powering AI with Ollama (local) or cloud agents like DeepSeek, using its Agentic RAG feature to make suggestions smarter and more relevant. It’s like giving your AI a roadmap to your code, so you can vibe code effortlessly.
CodeCompass: AI-powered Vibe Coding with MCP. Connects Git repositories to AI assistants like Claude, using Ollama for privacy or OpenAI for cloud. Integrates with VSCode, Cursor, and more.
星标
14
工具
—
可用率
—
Features
Codebase Analysis: Maps your repository's structure and dependencies, now with support for indexing very large files through automated chunking.
Smart AI Context with Agentic RAG: Utilizes a sophisticated Retrieval Augmented Generation (RAG) approach. The central agent_query tool intelligently orchestrates internal capabilities to gather comprehensive context. This includes analyzing git diff information (summarized if large), dynamically summarizing extensive file lists or code snippets, and more, ensuring AI suggestions are highly relevant.
Intelligent Agent Orchestration: The core agent_query tool allows the AI to plan and execute multi-step tasks. It can proactively use a suite of internal capabilities to:
Search code (capability_searchCodeSnippets)
Retrieve full file content (capability_getFullFileContent), with summarization for large files.
List directory contents (capability_listDirectory).
Analyze repository overviews including diffs and relevant snippets (capability_getRepositoryOverview).
Request more search results (capability_fetchMoreSearchResults) or more processing time if a query is complex.
Flexible Setup: Runs locally with Ollama or connects to cloud AI like DeepSeek.
Highly Configurable: Offers extensive environment variables to fine-tune indexing parameters, agent behavior (like loop steps and refinement iterations), context processing limits, and specific LLM models for tasks like summarization.
Project Status and Roadmap
Current Status:
CodeCompass has successfully implemented its core features, including:
Codebase analysis using Qdrant Vector Store.
Agentic RAG (Retrieval Augmented Generation) for intelligent AI suggestions.
Flexible integration with local LLMs via Ollama (e.g., llama3.1:8b, nomic-embed-text:v1.5) and cloud-based LLMs like DeepSeek.
The project is actively maintained and considered stable for its current feature set.
Future Enhancements (Under Consideration):
While the core functionality is robust, potential future directions include:
Support for a broader range of LLM providers (e.g., OpenAI, Gemini, Claude).
More sophisticated agent capabilities and additional tool integrations.
Enhanced repository indexing techniques for even more precise context retrieval.
Streamlined user configuration and an even smoother setup experience.
Deeper integrations with various IDEs and development workflows.
We welcome community contributions and suggestions for future development! Please see our CONTRIBUTING.md.
This installs CodeCompass globally. You can then run it from any directory.
Usage
CodeCompass can be run in two main modes: as a server (default) or as a client to execute specific tools against a running server.
1. Running the CodeCompass Server:
To start the CodeCompass server, navigate to the root of your git repository and run:
codecompass [repoPath] [--port <number>]
[repoPath] (optional): Path to the repository you want CodeCompass to analyze. If omitted, it defaults to the current directory (.).
--port <number> (optional): Specify the HTTP port for the server. This overrides the HTTP_PORT environment variable and the default port (3001).
Examples:
# Start server for the current directory on the default port
codecompass
# Start server for a specific repository
codecompass /path/to/your/project
# Start server for the current directory on port 3005
codecompass --port 3005
Once started, CodeCompass will begin indexing your repository (if it hasn't already). MCP communication is primarily handled via stdio. A utility HTTP server also runs on the configured port (default 3001) for health checks, indexing status, and to receive repository update notifications from Git hooks.
Port Conflicts for Utility HTTP Server:
If the configured utility HTTP port is in use by another CodeCompass instance, the new instance will disable its own utility HTTP server and run with stdio MCP. Its utility-related MCP tools (like get_indexing_status, trigger_repository_update) will automatically relay requests to the existing instance's HTTP utility endpoints. If the port is taken by a non-CodeCompass service, the new instance will log an error and exit. Git hooks should generally target the HTTP port of the primary running CodeCompass instance.
2. Executing Tools via CLI Client Mode:
To execute specific tools, CodeCompass CLI spawns a dedicated server instance and communicates with it via stdio using the Model Context Protocol (MCP).
<tool_name>: The name of the tool to execute (see list below).
[json_parameters] (optional): A JSON string containing parameters for the tool. For tools that take no parameters, this can be omitted or an empty JSON object {} can be used.
--repo <path> or -r <path> (optional): Specify the repository path for the tool's context. Defaults to the current directory (.). The spawned server will use this repository.
--json or -j (optional): Output the raw JSON response from the tool. Useful for scripting or debugging.
--port <number> (optional): Specify the utility HTTP port for the spawned server instance. This is primarily for consistency if the spawned server needs to interact with other instances or if its utility port needs to be specific, though core MCP communication for the tool is via stdio.
Available Tools for CLI Client Execution:
agent_query <json_params>: Orchestrates capabilities to answer complex queries.
Example: codecompass agent_query '{"query": "How is user authentication handled?", "sessionId": "my-session-123"}'
With JSON output: codecompass agent_query '{"query": "Show me user models"}' --json
Example: codecompass generate_suggestion '{"query": "optimize this database query"}'
get_repository_context <json_params>: Provides a high-level summary of repository context.
Example: codecompass get_repository_context '{"query": "main API components"}'
General CLI Commands:
codecompass --help or codecompass -h: Show help message.
codecompass --version or codecompass -v: Show CodeCompass version.
codecompass changelog [--verbose]: Show the project changelog.
With CodeCompass set up, use natural language prompts in Cursor or other AI tools to vibe code—interact with your codebase intuitively. The Agentic RAG feature, powered by Qdrant and Ollama/DeepSeek, ensures your AI understands your code’s context for precise results. The CLI client mode offers another way to interact with CodeCompass tools directly.
Configuration
CodeCompass relies on environment variables for its configuration. These variables can be set in several ways:
Directly in your shell: For the current session or persistently by adding them to your shell's profile script.
System-wide: Setting them at the operating system level.
Through MCP client settings: If you're using CodeCompass via an MCP client like Cursor or Cline, you can often define environment variables within their respective configuration files (e.g., mcp.json for Cursor). This is detailed in the "Setting Up with Cursor" section.
Using a .env file: For convenience, especially during local development, you can place a .env file in the root directory of the repository you are analyzing with CodeCompass. CodeCompass will load variables from this file.
Below are instructions for setting environment variables directly in your shell or system-wide, followed by a list of common variables.
Setting Environment Variables
For Linux and macOS:
You can set an environment variable for the current terminal session using the export command:
Here's a list of important environment variables. If you choose to use a .env file, you can copy this structure.
# --- General Configuration ---
# LOG_LEVEL: Logging level (e.g., error, warn, info, verbose, debug, silly). Default: info
# LOG_LEVEL=info
# --- Qdrant Configuration ---
# QDRANT_HOST: URL for the Qdrant vector store server.
QDRANT_HOST=http://localhost:6333
# COLLECTION_NAME: Name of the Qdrant collection for this repository.
# It's good practice to use a unique name per repository if you manage multiple.
# COLLECTION_NAME=codecompass_default_collection
# QDRANT_SEARCH_LIMIT_DEFAULT: Default number of results to fetch from Qdrant during standard searches. Default: 10
# QDRANT_SEARCH_LIMIT_DEFAULT=10
# --- Ollama Configuration (for local LLM and embeddings) ---
# OLLAMA_HOST: URL for the Ollama server.
OLLAMA_HOST=http://localhost:11434
# --- LLM Provider Configuration ---
# LLM_PROVIDER: Specifies the primary LLM provider for generating suggestions.
# Supported values: "ollama", "deepseek", "openai", "gemini", "claude". Default: "ollama"
LLM_PROVIDER=ollama
# SUGGESTION_MODEL: The specific model to use for suggestions.
# If LLM_PROVIDER="ollama", example: "llama3.1:8b", "codellama:7b"
# If LLM_PROVIDER="deepseek", example: "deepseek-coder"
# If LLM_PROVIDER="openai", example: "gpt-4-turbo-preview", "gpt-3.5-turbo"
# If LLM_PROVIDER="gemini", example: "gemini-pro"
# If LLM_PROVIDER="claude", example: "claude-2", "claude-3-opus-20240229"
# Default for Ollama: "llama3.1:8b"
SUGGESTION_MODEL=llama3.1:8b
# EMBEDDING_PROVIDER: Specifies the provider for generating embeddings.
# Currently, "ollama" is the primary supported embedding provider. Default: "ollama"
EMBEDDING_PROVIDER=ollama
# EMBEDDING_MODEL: The specific model to use for embeddings via Ollama.
# Default: "nomic-embed-text:v1.5"
EMBEDDING_MODEL=nomic-embed-text:v1.5
# --- Cloud Provider API Keys (only needed if using respective providers) ---
# DEEPSEEK_API_KEY: Your API key for DeepSeek.
# DEEPSEEK_API_KEY=your_deepseek_api_key_here
# OPENAI_API_KEY: Your API key for OpenAI.
# OPENAI_API_KEY=your_openai_api_key_here
# GEMINI_API_KEY: Your API key for Google Gemini.
# GEMINI_API_KEY=your_gemini_api_key_here
# CLAUDE_API_KEY: Your API key for Anthropic Claude.
# CLAUDE_API_KEY=your_claude_api_key_here
# --- DeepSeek Specific (Optional) ---
# DEEPSEEK_API_URL: Custom API URL for DeepSeek if not using the default.
# Default: "https://api.deepseek.com/chat/completions"
# DEEPSEEK_API_URL=https://api.deepseek.com/chat/completions
# DEEPSEEK_RPM_LIMIT: Requests per minute limit for DeepSeek. Default: 20
# DEEPSEEK_RPM_LIMIT=20
# --- Agent Configuration ---
# MAX_FILES_FOR_SUGGESTION_CONTEXT_NO_SUMMARY: Maximum number of files to list directly in the generate_suggestion tool's context
# before attempting to summarize the file list using an LLM. Default: 15
# MAX_FILES_FOR_SUGGESTION_CONTEXT_NO_SUMMARY=15
# MAX_SNIPPET_LENGTH_FOR_CONTEXT_NO_SUMMARY: Maximum length of a code snippet to include in context without summarization.
# Snippets longer than this will be summarized by an LLM if available. Default: 1500
# MAX_SNIPPET_LENGTH_FOR_CONTEXT_NO_SUMMARY=1500
# MAX_DIFF_LENGTH_FOR_CONTEXT_TOOL: Maximum length of a git diff to include in context without summarization.
# Diffs longer than this will be summarized by an LLM if available. Default: 3000
# MAX_DIFF_LENGTH_FOR_CONTEXT_TOOL=3000
# AGENT_DEFAULT_MAX_STEPS: Default maximum number of steps (tool calls or LLM responses) the agent will take. Default: 10
# AGENT_DEFAULT_MAX_STEPS=10
# AGENT_ABSOLUTE_MAX_STEPS: Absolute maximum number of steps the agent can take, even if it requests more. Default: 15
# AGENT_ABSOLUTE_MAX_STEPS=15
# MAX_REFINEMENT_ITERATIONS: Maximum number of iterations for query refinement. Default: 3
# MAX_REFINEMENT_ITERATIONS=3
# FILE_INDEXING_CHUNK_SIZE_CHARS: Target size for chunks when indexing large files (in characters). Default: 1000
# FILE_INDEXING_CHUNK_SIZE_CHARS=1000
# FILE_INDEXING_CHUNK_OVERLAP_CHARS: Overlap between chunks when indexing large files (in characters). Default: 200
# FILE_INDEXING_CHUNK_OVERLAP_CHARS=200
# SUMMARIZATION_MODEL: LLM model to use for summarization tasks (e.g., long diffs, long snippets).
# If empty or not set, defaults to the SUGGESTION_MODEL.
# SUMMARIZATION_MODEL= # Example: llama3.1:8b or deepseek-coder
# REFINEMENT_MODEL: LLM model to use for LLM-powered query refinement.
# If empty or not set, defaults to the SUGGESTION_MODEL.
# REFINEMENT_MODEL= # Example: llama3.1:8b or deepseek-coder
# REQUEST_ADDITIONAL_CONTEXT_MAX_SEARCH_RESULTS: Number of search results to fetch when agent uses 'request_additional_context' with type 'MORE_SEARCH_RESULTS'. Default: 20
# REQUEST_ADDITIONAL_CONTEXT_MAX_SEARCH_RESULTS=20
# AGENT_QUERY_TIMEOUT: Timeout in milliseconds for agent's LLM reasoning steps. Default: 180000 (3 minutes)
# AGENT_QUERY_TIMEOUT=180000
Note: When setting environment variables directly or via MCP client configurations, you do not need to create a .env file. The list above serves as a reference for the variable names and their purposes. For a local setup with Ollama, the default settings often work without needing to set many environment variables, unless you want to customize models or providers. If using cloud providers like DeepSeek, setting the respective API_KEY and adjusting LLM_PROVIDER and SUGGESTION_MODEL is necessary.
The HTTP port can also be set via the --port <number> CLI argument when starting the server, which takes precedence over the HTTP_PORT environment variable and defaults.
Troubleshooting
Model Not Found (Ollama):
Ensure you have pulled the correct model names using ollama pull <model_name>.
Verify installed models with ollama list.
Check that your OLLAMA_HOST environment variable is correct and the Ollama server is accessible.
Ensure the SUGGESTION_MODEL (if using Ollama for suggestions) and EMBEDDING_MODEL environment variables match the models you have pulled.
Connection Refused (Ollama/Qdrant):
Verify the Ollama server is running (e.g., ollama serve or the Ollama application).
Verify the Qdrant Docker container is running (docker ps) and accessible on the configured host/port (default http://localhost:6333).
Check your firewall settings if they might be blocking local connections.
API Key Issues (DeepSeek, OpenAI, Gemini, Claude):
Double-check that the respective API key (e.g., DEEPSEEK_API_KEY) is correctly set in your environment variables or .env file.
Ensure the key is valid and has not expired or reached its quota.
For DeepSeek, you can use the npm run test-deepseek script (if available in your setup) to diagnose connection issues.
Incorrect Provider/Model Mismatch:
Ensure LLM_PROVIDER and SUGGESTION_MODEL environment variables are compatible (e.g., use a DeepSeek model name like deepseek-coder when LLM_PROVIDER=deepseek).
General Issues:
Check CodeCompass logs for more detailed error messages. You might need to set LOG_LEVEL=debug for more verbose output.
Ensure Node.js and Docker are correctly installed and running.
Replace your_deepseek_api_key with your DeepSeek API key (or omit for Ollama). You can set any of the environment variables listed in the "Configuration" section (e.g., LLM_PROVIDER, SUGGESTION_MODEL) within this env block, and CodeCompass will use them.
Restart Cursor.
Note: For Cline in VSCode, configure similarly in cline_mcp_settings.json (see Cline Docs). Environment variables set in this manner will also be recognized by CodeCompass.
Contributing
Fork, branch, and submit a pull request. See CONTRIBUTING.md.
CodeCompass: AI-powered Vibe Coding with MCP. Connects Git repositories to AI assistants like Claude, using Ollama for privacy or OpenAI for cloud. Integrates with VSCode, Cursor, and more.
上次检查时间：2026-10-28。 CodeCompass: AI-powered Vibe Coding with MCP. Connects Git repositories to AI assistants like Claude, using Ollama for privacy or OpenAI for cloud. Integrates with VSCode, Cursor, and more.
