context-portal

Context Portal MCP (ConPort)

(It's a memory bank!)

           

A database-backed Model Context Protocol (MCP) server for managing structured project context, designed to be used by AI assistants and developer tools within IDEs and other interfaces.

What is Context Portal MCP server (ConPort)?

Context Portal (ConPort) is your project's memory bank. It's a tool that helps AI assistants understand your specific software project better by storing important information like decisions, tasks, and architectural patterns in a structured way. Think of it as building a project-specific knowledge base that the AI can easily access and use to give you more accurate and helpful responses.

What it does:

• Keeps track of project decisions, progress, and system designs.
• Stores custom project data (like glossaries or specs).
• Helps AI find relevant project information quickly (like a smart search).
• Enables AI to use project context for better responses (RAG).
• More efficient for managing, searching, and updating context compared to simple text file-based memory banks.

ConPort provides a robust and structured way for AI assistants to store, retrieve, and manage various types of project context. It effectively builds a project-specific knowledge graph, capturing entities like decisions, progress, and architecture, along with their relationships. This structured knowledge base, enhanced by vector embeddings for semantic search, then serves as a powerful backend for Retrieval Augmented Generation (RAG), enabling AI assistants to access precise, up-to-date information for more context-aware and accurate responses.

It replaces older file-based context management systems by offering a more reliable and queryable database backend (SQLite per workspace). ConPort is designed to be a generic context backend, compatible with various IDEs and client interfaces that support MCP.

Key features include:

• Structured context storage using SQLite (one DB per workspace, automatically created).
• MCP server (context_portal_mcp) built with Python/FastAPI.
• A comprehensive suite of defined MCP tools for interaction (see "Available ConPort Tools" below).
• Multi-workspace support via workspace_id.
• Primary deployment mode: STDIO for tight IDE integration.
• Enables building a dynamic project knowledge graph with explicit relationships between context items.
• Includes vector data storage and semantic search capabilities to power advanced RAG.
• Serves as an ideal backend for Retrieval Augmented Generation (RAG), providing AI with precise, queryable project memory.
• Provides structured context that AI assistants can leverage for prompt caching with compatible LLM providers.

Prerequisites

Before you begin, ensure you have the following installed:

• Python: Version 3.8 or higher is recommended.
  • Download Python
  • Ensure Python is added to your system's PATH during installation (especially on Windows).
• uv: (Highly Recommended) A fast Python environment and package manager. Using uv significantly simplifies virtual environment creation and dependency installation.
  • Install uv
  • If you choose not to use uv, you can use standard Python venv and pip, but uv is preferred for this project.

Installation via PyPI:

Create and activate a virtual environment in the directory where you install your MCP servers:

Using uv (recommended):

Activate the environment:

Linux/macOS (bash/zsh):

Windows (Command Prompt):

Windows (PowerShell):

(If you encounter execution policy issues in PowerShell, you might need to run Set-ExecutionPolicy RemoteSigned -Scope CurrentUser first.)

Using standard venv (if not using uv):

In your MCP server directory:

Activation commands are the same as for uv above.

Install ConPort via PyPi:

The PyPI installation command for context-portal-mcp using uv is:

If you are using standard pip within a virtual environment, the command is:

Configuration for PyPI Installation

If you installed ConPort via PyPI (pip install context-portal-mcp), the ConPort server can be launched directly using the Python interpreter within your virtual environment. This method is generally more reliable as it explicitly points to the executable.

Important: You MUST replace the placeholder path /home/USER/PATH/TO/.venv/bin/python (or C:\\Users\\USER\\PATH\\TO\\.venv\\Scripts\\python.exe on Windows) with the absolute path to the Python executable within your specific ConPort virtual environment.

Linux/macOS Example:

Windows Example:

• command: This must be the absolute path to the python (or python.exe on Windows) executable within the .venv of your ConPort installation.
• args: Contains the arguments to run the ConPort server module (-m context_portal_mcp.main) and the server's arguments (--mode stdio --workspace_id "${workspaceFolder}").
• ${workspaceFolder}: This IDE variable is used to automatically provide the absolute path of the current project workspace.
• --log-file: Optional: Path to a file where server logs will be written. If not provided, logs are directed to stderr (console). Useful for persistent logging and debugging server behavior.
• --log-level: Optional: Sets the minimum logging level for the server. Valid choices are DEBUG, INFO, WARNING, ERROR, CRITICAL. Defaults to INFO. Set to DEBUG for verbose output during development or troubleshooting.

When installed via PyPI, the conport-mcp command is available directly in your virtual environment's PATH. The command to run the server simplifies to:

Installation from Git Repository

These instructions guide you through setting up ConPort by cloning its Git repository and installing dependencies. Using a virtual environment is crucial.

1. Clone the Repository: Open your terminal or command prompt and run:
   git clone https://github.com/GreatScottyMac/context-portal.git cd context-portal
2. Create and Activate a Virtual Environment:
   • Using uv (recommended): In the context-portal directory:
     uv venv
     • Activate the environment:
       • Linux/macOS (bash/zsh):
         source .venv/bin/activate
       • Windows (Command Prompt):
         .venv\Scripts\activate.bat
       • Windows (PowerShell):
         .venv\Scripts\Activate.ps1
         (If you encounter execution policy issues in PowerShell, you might need to run Set-ExecutionPolicy RemoteSigned -Scope CurrentUser first.)
   • Using standard venv (if not using uv): In the context-portal directory:
     python3 -m venv .venv # Or 'python -m venv .venv'
     • Activation commands are the same as for uv above.
3. Install Dependencies: With your virtual environment activated:
   • Using uv (recommended):
     uv pip install -r requirements.txt
     Note: uv can often detect and use the .venv in the current directory even without explicit activation for uv pip install commands. However, activation is still good practice, especially if you intend to run Python scripts directly.
   • Using standard pip:
     pip install -r requirements.txt
4. Verify Installation (Optional): Ensure your virtual environment is activated.
   • Using uv:
     uv run python src/context_portal_mcp/main.py --help
   • Using standard python: bash python src/context_portal_mcp/main.py --help This should output the command-line help for the ConPort server.

Recommended Configuration (Direct Python Invocation):

This configuration directly invokes the Python interpreter from the context-portal virtual environment. It's a reliable method that does not depend on uv being the command or the client supporting a cwd field for the server process.

Important:

• You MUST replace placeholder paths with the absolute paths corresponding to where you have cloned and set up your context-portal repository.
• The "${workspaceFolder}" variable for the --workspace_id argument is a common IDE placeholder that should expand to the absolute path of your current project workspace.

Linux/macOS Example:

Imagine your context-portal repository is cloned at /home/youruser/mcp-servers/context-portal.

Windows Example:

Imagine your context-portal repository is cloned at C:\Users\YourUser\MCP-servers\context-portal. Note the use of double backslashes \\ for paths in JSON strings.

• command: This must be the absolute path to the python (or python.exe on Windows) executable within the .venv of your context-portal installation.
• First argument in args: This must be the absolute path to the main.py script within your context-portal installation.
• --workspace_id "${workspaceFolder}": This tells ConPort which project's context to manage. ${workspaceFolder} should be resolved by your IDE to the current project's root path.
• --log-file: Optional: Path to a file where server logs will be written. If not provided, logs are directed to stderr (console). Useful for persistent logging and debugging server behavior.
• --log-level: Optional: Sets the minimum logging level for the server. Valid choices are DEBUG, INFO, WARNING, ERROR, CRITICAL. Defaults to INFO. Set to DEBUG for verbose output during development or troubleshooting.

When installed via cloned Git repository, the IDE will typically construct and run a command similar to this:

/path/to/your/context-portal/ is the absolute path where you cloned the context-portal repository. "/actual/path/to/your/project_workspace" is the absolute path to the root of the project whose context ConPort will manage (e.g., ${workspaceFolder} in VS Code). ConPort automatically creates its database at your_project_workspace/context_portal/context.db.

Purpose of the --workspace_id Command-Line Argument:

When you launch the ConPort server, particularly in STDIO mode (--mode stdio), the --workspace_id argument serves several key purposes:

1. Initial Server Context: It provides the server process with the absolute path to the project workspace it should initially be associated with.
2. Critical Safety Check: In STDIO mode, this path is used to perform a vital check that prevents the server from mistakenly creating its database files (context.db, conport_vector_data/) inside its own installation directory. This protects against misconfigurations where the client might not correctly provide the workspace path.
3. Client Launch Signal: It's the standard way for an MCP client (like an IDE extension) to signal to the server which project it is launching for.

Important Note: The --workspace_id provided at server startup is not automatically used as the workspace_id parameter for every subsequent MCP tool call. ConPort tools are designed to require the workspace_id parameter explicitly in each call (e.g., get_product_context({"workspace_id": "..."})). This design supports the possibility of a single server instance managing multiple workspaces and ensures clarity for each operation. Your client IDE/MCP client is responsible for providing the correct workspace_id with each tool call.

Key Takeaway: ConPort critically relies on an accurate --workspace_id to identify the target project. Ensure this argument correctly resolves to the absolute path of your project workspace, either through IDE variables like ${workspaceFolder} or by providing a direct absolute path.

Clearing Python Bytecode Cache

Sometimes, after updating or reinstalling ConPort, you might encounter unexpected behavior or errors due to stale Python bytecode files (.pyc) stored in __pycache__ directories. Clearing this cache can resolve such issues.

You can use the find command on Unix-like systems (Linux, macOS, WSL) to locate and remove these files and directories.

1. Remove __pycache__ directories:
   find . -type d -name "__pycache__" -exec rm -rf {} +
2. Remove .pyc files:
   find . -type f -name "*.pyc" -delete

Where to run these commands:

The directory where you should run these commands depends on how you installed ConPort:

• If installed from the Git repository: Run these commands from the root directory of your cloned context-portal repository.
• If installed via PyPI: Run these commands from within the site-packages directory of the Python environment where ConPort is installed (e.g., from the root of your virtual environment's lib/pythonX.Y/site-packages/).
• If running from the Docker image: You would typically run these commands inside the running Docker container using docker exec <container_id> <command>.

Usage with LLM Agents (Custom Instructions)

ConPort's effectiveness with LLM agents is significantly enhanced by providing specific custom instructions or system prompts to the LLM. This repository includes tailored strategy files for different environments:

• For Roo Code:
  • roo_code_conport_strategy: Contains detailed instructions for LLMs operating within the Roo Code VS Code extension, guiding them on how to use ConPort tools for context management.
• For CLine:
  • cline_conport_strategy: Contains detailed instructions for LLMs operating within the Cline VS Code extension, guiding them on how to use ConPort tools for context management.
• For Windsurf Cascade:
  • cascade_conport_strategy: Specific guidance for LLMs integrated with the Windsurf Cascade environment. Important: When initiating a session in Cascade, it is necessary to explicity tell the LLM:
  Initialize according to custom instructions
• For General/Platform-Agnostic Use:
  • generic_conport_strategy: Provides a platform-agnostic set of instructions for any MCP-capable LLM. It emphasizes using ConPort's get_conport_schema operation to dynamically discover the exact ConPort tool names and their parameters, guiding the LLM on when and why to perform conceptual interactions (like logging a decision or updating product context) rather than hardcoding specific tool invocation details.

How to Use These Strategy Files:

1. Identify the strategy file relevant to your LLM agent's environment.
2. Copy the entire content of that file.
3. Paste it into your LLM's custom instructions or system prompt area. The method varies by LLM platform (IDE extension settings, web UI, API configuration).

These instructions equip the LLM with the knowledge to:

• Initialize and load context from ConPort.
• Update ConPort with new information (decisions, progress, etc.).
• Manage custom data and relationships.
• Understand the importance of workspace_id. Important Tip for Starting Sessions: To ensure the LLM agent correctly initializes and loads context, especially in interfaces that might not always strictly adhere to custom instructions on the first message, it's a good practice to start your interaction with a clear directive like: Initialize according to custom instructions. This can help prompt the agent to perform its ConPort initialization sequence as defined in its strategy file.

Initial ConPort Usage in a Workspace

When you first start using ConPort in a new or existing project workspace, the ConPort database (context_portal/context.db) will be automatically created by the server if it doesn't exist. To help bootstrap the initial project context, especially the Product Context, consider the following:

Using a projectBrief.md File (Recommended)

1. Create projectBrief.md: In the root directory of your project workspace, create a file named projectBrief.md.
2. Add Content: Populate this file with a high-level overview of your project. This could include:
   • The main goal or purpose of the project.
   • Key features or components.
   • Target audience or users.
   • Overall architectural style or key technologies (if known).
   • Any other foundational information that defines the project.
3. Automatic Prompt for Import: When an LLM agent using one of the provided ConPort custom instruction sets (e.g., roo_code_conport_strategy) initializes in the workspace, it is designed to:
   • Check for the existence of projectBrief.md.
   • If found, it will read the file and ask you if you'd like to import its content into the ConPort Product Context.
   • If you agree, the content will be added to ConPort, providing an immediate baseline for the project's Product Context.

Manual Initialization

If projectBrief.md is not found, or if you choose not to import it:

• The LLM agent (guided by its custom instructions) will typically inform you that the ConPort Product Context appears uninitialized.
• It may offer to help you define the Product Context manually, potentially by listing other files in your workspace to gather relevant information.

By providing initial context, either through projectBrief.md or manual entry, you enable ConPort and the connected LLM agent to have a better foundational understanding of your project from the start.

Available ConPort Tools

The ConPort server exposes the following tools via MCP, allowing interaction with the underlying project knowledge graph. This includes tools for semantic search powered by vector data storage. These tools facilitate the Retrieval aspect crucial for Augmented Generation (RAG) by AI agents. All tools require a workspace_id argument (string, required) to specify the target project workspace.

• Product Context Management:
  • get_product_context: Retrieves the overall project goals, features, and architecture.
  • update_product_context: Updates the product context. Accepts full content (object) or patch_content (object) for partial updates (use __DELETE__ as a value in patch to remove a key).
• Active Context Management:
  • get_active_context: Retrieves the current working focus, recent changes, and open issues.
  • update_active_context: Updates the active context. Accepts full content (object) or patch_content (object) for partial updates (use __DELETE__ as a value in patch to remove a key).
• Decision Logging:
  • log_decision: Logs an architectural or implementation decision.
    • Args: summary (str, req), rationale (str, opt), implementation_details (str, opt), tags (list[str], opt).
  • get_decisions: Retrieves logged decisions.
    • Args: limit (int, opt), tags_filter_include_all (list[str], opt), tags_filter_include_any (list[str], opt).
  • search_decisions_fts: Full-text search across decision fields (summary, rationale, details, tags).
    • Args: query_term (str, req), limit (int, opt).
  • delete_decision_by_id: Deletes a decision by its ID.
    • Args: decision_id (int, req).
• Progress Tracking:
  • log_progress: Logs a progress entry or task status.
    • Args: status (str, req), description (str, req), parent_id (int, opt), linked_item_type (str, opt), linked_item_id (str, opt).
  • get_progress: Retrieves progress entries.
    • Args: status_filter (str, opt), parent_id_filter (int, opt), limit (int, opt).
  • update_progress: Updates an existing progress entry.
    • Args: progress_id (int, req), status (str, opt), description (str, opt), parent_id (int, opt).
  • delete_progress_by_id: Deletes a progress entry by its ID.
    • Args: progress_id (int, req).
• System Pattern Management:
  • log_system_pattern: Logs or updates a system/coding pattern.
    • Args: name (str, req), description (str, opt), tags (list[str], opt).
  • get_system_patterns: Retrieves system patterns.
    • Args: tags_filter_include_all (list[str], opt), tags_filter_include_any (list[str], opt).
  • delete_system_pattern_by_id: Deletes a system pattern by its ID.
    • Args: pattern_id (int, req).
• Custom Data Management:
  • log_custom_data: Stores/updates a custom key-value entry under a category. Value is JSON-serializable.
    • Args: category (str, req), key (str, req), value (any, req).
  • get_custom_data: Retrieves custom data.
    • Args: category (str, opt), key (str, opt).
  • delete_custom_data: Deletes a specific custom data entry.
    • Args: category (str, req), key (str, req).
  • search_project_glossary_fts: Full-text search within the 'ProjectGlossary' custom data category.
    • Args: query_term (str, req), limit (int, opt).
  • search_custom_data_value_fts: Full-text search across all custom data values, categories, and keys.
    • Args: query_term (str, req), category_filter (str, opt), limit (int, opt).
• Context Linking:
  • link_conport_items: Creates a relationship link between two ConPort items, explicitly building out the project knowledge graph.
    • Args: source_item_type (str, req), source_item_id (str, req), target_item_type (str, req), target_item_id (str, req), relationship_type (str, req), description (str, opt).
  • get_linked_items: Retrieves items linked to a specific item.
    • Args: item_type (str, req), item_id (str, req), relationship_type_filter (str, opt), linked_item_type_filter (str, opt), limit (int, opt).
• History & Meta Tools:
  • get_item_history: Retrieves version history for Product or Active Context.
    • Args: item_type ("product_context" | "active_context", req), version (int, opt), before_timestamp (datetime, opt), after_timestamp (datetime, opt), limit (int, opt).
  • get_recent_activity_summary: Provides a summary of recent ConPort activity.
    • Args: hours_ago (int, opt), since_timestamp (datetime, opt), limit_per_type (int, opt, default: 5).
  • get_conport_schema: Retrieves the schema of available ConPort tools and their arguments.
• Import/Export:
  • export_conport_to_markdown: Exports ConPort data to markdown files.
    • Args: output_path (str, opt, default: "./conport_export/").
  • import_markdown_to_conport: Imports data from markdown files into ConPort.
    • Args: input_path (str, opt, default: "./conport_export/").
• Batch Operations:
  • batch_log_items: Logs multiple items of the same type (e.g., decisions, progress entries) in a single call.
    • Args: item_type (str, req - e.g., "decision", "progress_entry"), items (list[dict], req - list of Pydantic model dicts for the item type).

Prompt Caching Strategy

ConPort can be used to provide structured context (including vector data for semantic search) that AI assistants can leverage for prompt caching with compatible LLM providers (like Google Gemini, Anthropic Claude, and OpenAI). Prompt caching reduces token costs and latency by reusing frequently used parts of prompts.

This repository includes a detailed strategy file (context_portal/prompt_caching_strategy.yml) that defines how an LLM assistant should identify cacheable content from ConPort and structure prompts for different providers.

Key aspects of the strategy include:

• Identifying Cacheable Content: Prioritizing large, stable context like Product Context, detailed System Patterns, or specific Custom Data entries (especially those flagged with a cache_hint: true metadata).
• Provider-Specific Interaction:
  • Implicit Caching (Gemini, OpenAI): Structure prompts by placing cacheable ConPort content at the absolute beginning of the prompt. The LLM provider automatically handles caching.
  • Explicit Caching (Anthropic): Insert a cache_control breakpoint after the cacheable ConPort content within the prompt payload.
• User Hints: ConPort's Custom Data can include metadata like cache_hint: true to explicitly guide the LLM assistant on content prioritization for caching.
• LLM Assistant Notification: The LLM assistant is instructed to notify the user when it structures a prompt for potential caching (e.g., [INFO: Structuring prompt for caching]).

By using ConPort to manage your project's knowledge and providing the LLM assistant with this prompt caching strategy, you can enhance the efficiency and cost-effectiveness of your AI interactions.

Further Reading

For a more in-depth understanding of ConPort's design, architecture, and advanced usage patterns, please refer to:

• conport_mcp_deep_dive.md

Contributing

Details on contributing to the ConPort project will be added here in the future.

License

This project is licensed under the Apache-2.0 license.