MCP Server

How external agents interact with the wiki-base via the Model Context Protocol: read-only by default.

The Idea

Every community, product, and service will eventually have an externalised knowledge base that agents can query. The MCP server is how external agents: Claude Code, Cursor, Windsurf, or any MCP-compatible client: interact with the Latent Space wiki-base programmatically.

Instead of humans reading docs or searching a website, their agents connect to the knowledge graph directly, search it, and bring back what's relevant.

Architecture

The standalone MCP server (apps/mcp-server-standalone/) connects directly to Turso via @libsql/client. It runs as a stdio process: the MCP client spawns it as a subprocess.

AI Agent (Claude Code / Cursor / Windsurf)
    ↓ stdio
latent-space-hub-mcp (Node.js process)
    ↓ libsql
Turso cloud SQLite

Published to NPM as latent-space-hub-mcp.

Quick Start

Add to your Claude Code, Cursor, or Windsurf MCP config:

{
  "mcpServers": {
    "latent-space-hub": {
      "command": "npx",
      "args": ["-y", "latent-space-hub-mcp"],
      "env": {
        "TURSO_DATABASE_URL": "your-turso-url",
        "TURSO_AUTH_TOKEN": "your-turso-token"
      }
    }
  }
}

Security

The MCP server is read-only by default. External agents can search and read the wiki-base but cannot create, update, or delete anything.

Write tools (node creation, edge creation, dimension management) are only registered when MCP_ALLOW_WRITES=true is explicitly set. This is intended for trusted internal use only: not for public-facing agent access.

Variable	Description	Default
`TURSO_DATABASE_URL`	Turso database URL (required)	:
`TURSO_AUTH_TOKEN`	Turso auth token (required)	:
`MCP_ALLOW_WRITES`	Enable write tools	`false`
`OPENAI_API_KEY`	Enables vector search in `ls_search_nodes`	:
`LSH_SKILLS_DIR`	Custom skills directory	`~/.latent-space-hub/skills/`

The Turso credentials are the primary security boundary. Do not share them publicly.

Read Tools (Always Available)

Tool	Description
`ls_get_context`	Graph overview: stats, top nodes, dimensions, available skills
`ls_search_nodes`	Hybrid search across nodes by title/description. Supports `node_type`, date range, and dimension filters
`ls_search_content`	Full-text + vector search through source material (transcripts, articles). Hybrid RRF fusion
`ls_get_nodes`	Load full node records by ID (up to 10 per call)
`ls_query_edges`	Find all connections for a specific node
`ls_list_dimensions`	List all dimensions/tags with node counts
`ls_sqlite_query`	Run read-only SQL: `SELECT`, `WITH`, `PRAGMA` only
`ls_list_skills`	List system and custom skills
`ls_read_skill`	Read a skill by name

Write Tools (Requires `MCP_ALLOW_WRITES=true`)

Tool	Description
`ls_add_node`	Create a new node (title + dimensions required)
`ls_update_node`	Update a node: content appends to notes, dimensions replace
`ls_create_edge`	Connect two nodes with a directional relationship and explanation
`ls_update_edge`	Update an edge's explanation
`ls_create_dimension`	Create a new dimension/tag
`ls_update_dimension`	Rename or update a dimension
`ls_delete_dimension`	Delete a dimension and remove from all nodes
`ls_write_skill`	Create or update a custom skill
`ls_delete_skill`	Delete a custom skill

Setup Options

Option	Config	Notes
NPX (standalone)	`"command": "npx", "args": ["-y", "latent-space-hub-mcp"]`	No local repo needed. Connects directly to Turso. Read-only by default.
Local stdio	`"command": "node", "args": ["path/to/apps/mcp-server/stdio-server.js"]`	Requires hub running (`npm run dev`). Proxies to Next.js API.
HTTP transport	`"url": "http://127.0.0.1:44145/mcp"`	For agents supporting HTTP MCP transport.

Slop Does Not Use MCP

The Discord bot (Slop) used to connect via MCP but has been decoupled. Slop now queries Turso directly with its own internal tools (slop_* prefix) in apps/bots/slop/. The MCP server is exclusively for external agent access.