DollhouseMCP is a free, open-source MCP (Model Context Protocol) server designed to enable AI customization through modular elements. It allows users to create, activate, edit, and manage multiple types of elements, including persona, skill, template, agent, memory, and ensemble. All elements are stored as readable Markdown or YAML files in a local portfolio, and users fully own and control these files.
Main capabilities:
- Element management: Supports creating, activating, searching, and browsing elements. Users can describe their needs in natural language, and the AI will generate the corresponding element.
- Local portfolio: Elements are stored in a local folder (on macOS/Linux, ~/.dollhouse/portfolio/; on Windows, %USERPROFILE%.dollhouse\portfolio), with 37 built-in starter elements, including the dollhouse-expert-suite ensemble.
- Community collections: Provides public GitHub collections, allowing users to browse and install elements created by other users, and to share their own elements.
- GitHub sync: Optionally backs up the portfolio to a GitHub repository, enabling cross-device synchronization and community sharing.
- Skill format conversion: Includes a built-in lossless bidirectional converter that can convert between Dollhouse Skills and the agent skills format adopted by Claude/Anthropic.
- MCP-AQL: Organizes all operations into 5 semantic endpoints (CRUDE: Create, Read, Update, Delete, Execute), reducing context token consumption and supporting progressive disclosure and introspection.
- Gatekeeper permission system: Server-side permission control, where elements can define allow, confirm, and deny policies, as well as external tool restrictions. Policies stack across all activated elements, with deny taking precedence.
- Agent execution: Every step of Dollhouse Agents passes through the MCP server, LLM, and Gatekeeper, including autonomy assessment, hard blocking of dangerous areas, and step logging.
- Other features: Web portfolio browser, batch operations, activationpersistence, universal backup, cache memory budget, NLP discovery, cross-element relationship mapping, security hardening, and more.
How to use:
- Installation: Supports an interactive setup wizard (one-click installation to Claude Desktop, Claude Code, Cursor, VS Code, Codex, Gemini CLI, Windsurf, Cline, LM Studio, etc.), or local installation via npm.
- Activate elements: Ask the AI in conversation to activate an element, for example, "Activate the Dollhouse code review persona."
- Create elements: Describe in natural language the type and content of the element you want to create.
- Browse the community: Browse and install community elements through the AI.
- Manage the portfolio: You can ask the AI to "open the portfolio browser" for visual management.
Use cases:
- Users who need to customize behavior, capabilities, and permissions for AI clients.
- Developers who want to reuse and manage AI configurations across different MCP clients.
- Teams that need fine-grained permission control (such as read-only mode and secure sandboxing).
- Users who want to share or obtain community AI elements.
Limitations:
- Advanced features (Gatekeeper confirmation flow, agent loop execution) have mainly been extensively tested on Claude Code and DollhouseMCP Bridge; other platforms may vary due to differences in the LLM's ability to interpret structured tool responses.
- Requires a Node.js 20+ environment.
- Some features (such as GitHub sync) require additional configuration.
- Specific fees, pricing, platform support details, etc. are not provided in the materials.
