Context Portal MCP

Install

Works in

• Roo Code
• Cline
• Windsurf
• Cursor
• SourceCraft
• Koda
• Coddy

npx skillfoxx add mcp/context-portal

What it does

Context Portal, or ConPort, is a Python MCP server that maintains a project memory bank for each workspace. A SQLite database is created automatically in a `context_portal/` folder, with a ChromaDB vector store beside it. The agent writes decisions with their rationale, task statuses, architecture patterns and free-form data such as a glossary, and later retrieves them by full-text and semantic search. Vectors are built by the local `all-MiniLM-L6-v2` model from the `sentence-transformers` package. Records are linked with the `link_conport_items` tool, and the links form a project knowledge graph. Product context and the current working context are stored with version history, and `get_item_history` returns earlier states. `get_recent_activity_summary` gathers recent changes, `batch_log_items` writes a batch of decisions or tasks in one call, and `export_conport_to_markdown` with `import_markdown_to_conport` move data to Markdown files and back. The database schema is updated through Alembic migrations. Every call carries a `workspace_id`, so one server can serve several projects. If no path is given at launch, ConPort finds the project root itself from `.git`, `package.json`, `pyproject.toml` and other markers. The repository contains agent instructions for Roo Code, Cline, Windsurf Cascade and a generic variant: they are pasted into the system prompt so the agent loads and updates memory by itself.

Who it is for. For developers who work with Roo Code, Cline, Windsurf or Cursor and want the agent to keep decisions and progress in a project database.

Example request

Log a decision in ConPort: we use Redis for the job queue because it is already in the project, and link it to the task about background email sending

MCP

Transport

stdio

Authentication

not required

Limitations

• The license is Apache-2.0. The manifest carries version 0.3.13 and the classifier Beta.
• The instructions say Python 3.8 or higher, but the manifest requires 3.10 or newer, so follow the manifest.
• Many editors do not expand `${workspaceFolder}` when launching the server. The author advises an absolute path in `--workspace_id`, or leaving it out at launch and passing it in every call. If an IDE passes the literal string `${workspaceFolder}`, ConPort ignores it and finds the root itself.
• Without agent instructions memory does not fill up: the server only answers calls. Windsurf Cascade needs a new session to be told directly to initialize according to custom instructions.
• When updating across versions with a database schema change the author advises reading `v0.2.4_UPDATE_GUIDE.md`: sometimes data has to be moved manually through export and import.

How to disable. Remove the `conport` entry from the MCP settings of the client (in the guide's example this is `mcp_settings.json`). Data stays in the workspace's `context_portal/` folder, which can be deleted by hand. Before deleting you can export memory to Markdown with the `export_conport_to_markdown` tool.

README in short

The front page explains what ConPort is and lists its features: SQLite per workspace, a vector store, a knowledge graph and Alembic migrations. Next come prerequisites, setup through `uvx` and a developer variant from a clone, a section on agent instructions, workspace auto-detection and a reference of about thirty tools with their arguments. The repository also holds a deep-dive document `conport_mcp_deep_dive.md`, an update guide, a Dockerfile, contribution rules and handler tests.