MCP

ThinkForge exposes your project memory through the Model Context Protocol (MCP), an open standard that lets AI assistants in compatible editors search, retrieve, and save content in your knowledge base.

When ThinkForge Desktop is running, MCP clients can access your project memory without any data leaving your machine. The MCP server runs locally alongside the desktop app.

What MCP enables

With MCP configured, AI assistants in your editor or tool can:

  • Search by meaning — hybrid search combining semantic similarity and keyword matching
  • Search by keywords — pure semantic search for content similar to a query
  • Search across all projects — find content in any ThinkForge project, not just the active one
  • Retrieve documents — get a specific document or content chunk by ID
  • Save artifacts — write generated notes, code, or documentation back into project memory
  • Search by tags — find content by assigned tags
  • List tags — see all tags in the current project
  • List projects — discover available ThinkForge projects
  • Get statistics — memory and index status information

Available tools

The ThinkForge MCP server provides these tools to connected clients:

  • hybrid_search — recommended default; combines meaning and keyword matching
  • semantic_search — pure vector similarity search
  • search_all_projects — cross-project search
  • list_projects — list project names for scoping
  • get_memory — retrieve a specific document or chunk
  • save_artifact — save generated content into memory
  • tag_search — find documents by tag
  • list_tags — list all project tags
  • get_statistics — index and memory statistics

Supported editors and tools

ThinkForge can auto-configure MCP for these editors from Settings → MCP Config:

  • Cursor
  • Windsurf
  • Kiro
  • VS Code (Copilot)
  • Cline

For editors that use stdio transport (like Claude Desktop), a separate adapter is available. See Claude Desktop for that setup.

Setup

  1. Open ThinkForge Desktop with a project active.
  2. Go to Settings → MCP Config.
  3. Select your editor from the list.
  4. ThinkForge writes the necessary configuration file for that editor.
  5. Restart your editor (or reload the AI assistant) to pick up the new MCP connection.

The MCP server runs on localhost (port 3001) and starts automatically with ThinkForge Desktop. No external setup, cloud services, or port forwarding required.

Agent behavior toggle

ThinkForge offers an optional Agent Behaviour setting that adds keyword triggers. When your IDE's AI assistant encounters phrases like "search memory" or "use think-forge," it is prompted to call the MCP tools before answering. This helps AI assistants proactively use your project knowledge without you explicitly asking.

Requirements

  • ThinkForge Desktop must be running
  • A project must be open (SK Memory needs an active project to search)
  • The editor must support MCP (most modern AI-enabled editors do)
Fully local. The MCP server runs on your machine. Search queries and results stay on localhost. Nothing is sent to external services through MCP — the AI assistant in your editor handles its own API calls separately.

Related