The three tools are deliberately few; the product is the UI in front of them, where you group skills and flip them on or off so an agent only carries what a task needs. The always-loaded versus dynamic split is the point — some knowledge should be present every session, some should be fetched by name only when relevant. One non-obvious step: the tools alone are not enough, the README wants `MCP_instructions.md` pasted into the agent's system prompt.
A self-hosted web application plus MCP server for managing AI skills, workflows and contexts. You organise skills into groups in a local web UI and toggle them on or off; the MCP server then loads what is enabled into an agent, either automatically or on demand.
- Every skill marked `always_loaded` pulled into context at once — `get_default_skills`
- All enabled skills listed with their descriptions and modes — `list_available_skills`
- A specific `dynamic` skill loaded on demand by name — `load_full_skill_context`
- A web UI to create, edit, organise and search skills and workflows, with per-group enable/disable toggles and drag-and-drop
- Two loading modes per skill — always-loaded, or dynamic on-demand
- Import from a local folder, a GitHub repository, or the Skills.sh registry, plus token counting to track context size
- Isolated workspaces (hubs), with the server aware of the active one
Node and Python. Install is a clone plus `chmod +x setup.sh` and `./setup.sh`, which generates `mcp_settings.json` with the correct paths; the web app runs with `npm start` on `http://localhost:3000`. The generated MCP entry launches `mcp_server.py` through the project's `.venv` Python. The package.json names the web app `context-manager-web` at 1.0.0. The README stresses one setup step beyond the tools: the contents of `MCP_instructions.md` must be copied into the agent's system prompt or rules, or the three tools have no usage guidance. A starter library of skills and workflows is copied into your hub on first setup.