A server for managing structured project context using SQLite, with support for vector embeddings for semantic search and Retrieval Augmented Generation (RAG).
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.
- 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.
- 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.
- Manages database schema evolution using**Alembic migrations**, ensuring seamless updates and data integrity.
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).
## Installation and Configuration (Recommended)
The recommended way to install and run ConPort is by using`uvx`to execute the package directly from PyPI. This method avoids the need to manually create and manage virtual environments.
### `uvx`Configuration (Recommended for most IDEs)
In your MCP client settings (e.g.,`mcp_settings.json`), use the following configuration:
```
`{ "mcpServers": { "conport": { "command": "uvx", "args": ](https://www.python.org/downloads/)[ "--from", "context-portal-mcp", "conport-mcp", "--mode", "stdio", "--workspace_id", "${workspaceFolder}", "--log-file", "./logs/conport.log", "--log-level", "INFO" ] } } }`
```
- **`command`**:`uvx`handles the environment for you.
- **`args`**: Contains the arguments to run the ConPort server.
- `${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.
Important: Many IDEs do not expand`${workspaceFolder}`when launching MCP servers. Use one of these safe options:
- Provide an absolute path for`--workspace_id`.
- Omit`--workspace_id`at launch and rely on per-call`workspace_id`(recommended if your client provides it on every call).
Alternative configuration (no`--workspace_id`at launch):
```
`{ "mcpServers": { "conport": { "command": "uvx", "args": [ "--from", "context-portal-mcp", "conport-mcp", "--mode", "stdio", "--log-file", "./logs/conport.log", "--log-level", "INFO" ] } } }`
```
If you omit`--workspace_id`, the server will skip pre-initialization and initialize the database on the first tool call using the`workspace_id`provided in that call.
## Installation for Developers (from Git Repository)
The most appropriate way to develop and test ConPort is to run it in your IDE as an MCP server using the configuration above. This exercises STDIO mode and real client behavior.
If you need to run against a local checkout and virtualenv, you can configure your MCP client to launch the dev server via`uv run`and your`.venv/bin/python`:
```
`{ "mcpServers": { "conport": { "command": "uv", "args": [ "run", "--python", ".venv/bin/python", "--directory", "<path to context-portal repo> ", "conport-mcp", "--mode", "stdio", "--log-file", "./logs/conport-dev.log", "--log-level", "DEBUG" ], "disabled": false } } }`
```
- Set`--directory`to your repo path; this uses your local checkout and venv interpreter.
- Logs go to`./logs/conport-dev.log`with`DEBUG`verbosity.
Set up for development or contribution via the Git repo.
```
`git clone https://github.com/GreatScottyMac/context-portal.git cd context-portal`
```
Activate it using your shell’s standard activation (e.g.,`source .venv/bin/activate`on macOS/Linux).
**Run in your IDE (recommended)**Configure your IDE’s MCP settings using the "uvx Configuration" or the dev`uv run`configuration shown above. This is the most representative test of ConPort in STDIO mode.
```
`uv run python src/context_portal_mcp/main.py --help`
```
- For`--workspace_id`behavior and IDE path handling, see the guidance under the "uvx Configuration" section above. Many IDEs do not expand`${workspaceFolder}`.
For pre-upgrade cleanup, including clearing Python bytecode cache, please refer to the[v0.2.4_UPDATE_GUIDE.md.
## 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:
- ](https://github.com/GreatScottyMac/context-portal/blob/HEAD/v0.2.4_UPDATE_GUIDE.md#1-pre-upgrade-cleanup)[`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.
- ](https://github.com/GreatScottyMac/context-portal/blob/main/conport-custom-instructions/roo_code_conport_strategy)[`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.
- ](https://github.com/GreatScottyMac/context-portal/blob/main/conport-custom-instructions/cline_conport_strategy)[`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`
```
- ](https://github.com/GreatScottyMac/context-portal/blob/main/conport-custom-instructions/cascade_conport_strategy)[`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.
- Identify the strategy file relevant to your LLM agent's environment.
- Copy the**entire content**of that file.
- 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.
### New Strategy Set: mem4sprint (What’s new)
The repository includes a new strategy/documentation set focused on sprint planning and operational flows:
- `conport-custom-instructions/mem4sprint.md`— concise guidance and patterns for using flat categories and valid FTS prefixes.
- `conport-custom-instructions/mem4sprint.schema_and_templates.md`— meta schema, compact starters, FTS query rules, and minimal operational call recipes.
- Flat category model (e.g.,`artifacts`,`rfc_doc`,`retrospective`,`ProjectGlossary`,`critical_settings`).
- Valid FTS5 prefixes only:`category:`,`key:`,`value_text:`for custom data;`summary:`,`rationale:`,`implementation_details:`,`tags:`for decisions.
- Handler-layer query normalization; database layer remains unchanged.
- Added mem4sprint strategy/docs with flattened categories and explicit FTS rules.
- Simplified examples and included minimal operational call recipes.
- Documentation clarifies IDE workspace path handling for MCP.
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)
- **Create`projectBrief.md`:**In the root directory of your project workspace, create a file named`projectBrief.md`.
- **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.
- 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.
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.
ConPort can automatically determine the correct`workspace_id`so you do not need to hardcode an absolute path in your MCP client configuration. This is especially useful for IDEs that fail to expand`${workspaceFolder}`when launching MCP servers.
Detection is enabled by default and can be controlled via CLI flags:
- `--auto-detect-workspace`(default: enabled) Turns on automatic detection.
- `--no-auto-detect`Disables detection (explicit`--workspace_id`or per-tool`workspace_id`must then be provided).
- `--workspace-search-start <path>`Optional starting directory for upward search (defaults to current working directory).
- Strong Indicators (fast path): Looks for high-confidence project roots containing any of:`package.json`,`.git`,`pyproject.toml`,`Cargo.toml`,`go.mod`,`pom.xml`.
- Multiple General Indicators: If ≥2 general indicators (README, license, build files, etc.) exist in a directory, it is treated as a root.
- Existing ConPort Workspace: Presence of a`context_portal/`directory indicates a valid workspace.
- MCP Environment Context: Honors environment variables like`VSCODE_WORKSPACE_FOLDER`or`CONPORT_WORKSPACE`when set and valid.
- Fallback: If no indicators are found, uses the starting directory verbosely (with a warning).
- `get_workspace_detection_info`(MCP tool) exposes a diagnostic dictionary showing:
- start_path
- detected_workspace
- detection_method (strong_indicators | multiple_indicators | existing_context_portal | fallback)
- indicators_found
- relevant environment variables
- Keep detection enabled unless you operate multi-root scenarios where explicit isolation per call is required.
- If an IDE passes the literal string`${workspaceFolder}`, ConPort will ignore it and auto-detect safely (logged at WARNING).
- For debugging ambiguous roots (e.g., nested repos), run the detection info tool to confirm which directory was selected.
Example MCP launch (relying fully on auto-detect):
```
`{ "mcpServers": { "conport": { "command": "uvx", "args": ](https://github.com/GreatScottyMac/context-portal/blob/main/conport-custom-instructions/generic_conport_strategy)[ "--from", "context-portal-mcp", "conport-mcp", "--mode", "stdio", "--log-level", "INFO" ] } } }`
```
To disable detection explicitly (forcing provided IDs only):
```
`{ "mcpServers": { "conport": { "command": "uvx", "args": [ "--from", "context-portal-mcp", "conport-mcp", "--mode", "stdio", "--no-auto-detect", "--workspace_id", "/absolute/path/to/project" ] } } }`
```
If you have a launcher that starts inside a deep subdirectory, provide a higher start path:
```
`conport-mcp --mode stdio --workspace-search-start ../../`
```
See`UNIVERSAL_WORKSPACE_DETECTION.md`for full rationale, edge cases, and troubleshooting.
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.
Note: For convenience, all integer-like parameters accept either numbers or digit-only strings (e.g., "10", " 3"). The server trims whitespace and coerces these to integers while preserving validation bounds (e.g., ge=1). Credit: @cipradu.
- **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).
- `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).
- `log_decision`: Logs an architectural or implementation decision.
- Args:`summary`(str, req),`rationale`(str, opt),`implementation_details`(str, opt),`tags`(list[str], opt).
- Args:`limit`(int, opt),`tags_filter_include_all`(list[str], opt),`tags_filter_include_any`(list[str], opt).
- Args:`query_term`(str, req),`limit`(int, opt).
- `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).
- Args:`status_filter`(str, opt),`parent_id_filter`(int, opt),`limit`(int, opt).
- Args:`progress_id`(int, req),`status`(str, opt),`description`(str, opt),`parent_id`(int, opt).
- `log_system_pattern`: Logs or updates a system/coding pattern.
- Args:`name`(str, req),`description`(str, opt),`tags`(list[str], opt).
