Infino MCP Server

io.github.infino-ai/mcp-server

Overview

Keyword, vector, hybrid, token/exact, and SQL retrieval over local or object-storage Infino data for AI agents.

Documentation

Overview

Infino MCP Server is a local stdio MCP server for keyword, semantic, hybrid, token/exact, and SQL retrieval over Infino data on local paths or object storage. It is published on npm as @infino-ai/mcp-server and registered as io.github.infino-ai/mcp-server.

Installation

MCP clients launch the package over stdio:

npx -y @infino-ai/mcp-server

Typical client config:

{
  "command": "npx",
  "args": ["-y", "@infino-ai/mcp-server"],
  "env": {
    "INFINO_MCP_URI": "/Users/me/.infino/memory"
  }
}

Claude Code setup from the README:

claude mcp add infino --scope user -e INFINO_MCP_URI=/Users/me/.infino/memory -- npx -y @infino-ai/mcp-server

The repository can also be installed as a Claude Code plugin with /plugin marketplace add infino-ai/infino-mcp and /plugin install infino@infino-ai.

Configuration

All runtime configuration is via environment variables. INFINO_MCP_URI points to a local path or bucket URI (s3://, gs://, az://) and defaults to ephemeral memory://. INFINO_MCP_ENABLE_WRITES enables add/update/delete tools and DDL/DML SQL when set to 1, true, or yes; otherwise the server is read-only. INFINO_MCP_EMBED_MODEL selects the Hugging Face feature-extraction model for local embeddings and defaults to Xenova/all-MiniLM-L6-v2. INFINO_MCP_S3_ENDPOINT and INFINO_MCP_S3_REGION configure non-AWS S3-compatible stores. Standard cloud credentials are used for S3, Google Cloud Storage, and Azure Blob.

Capabilities

Tools include infino_list_tables, infino_describe_table, infino_keyword_search, infino_semantic_search, infino_hybrid_search, infino_token_match, infino_exact_match, and infino_sql. When writes are enabled, infino_add_documents, infino_update_documents, and infino_delete_documents are also advertised. Semantic and hybrid search use local embeddings; no embedding API key is required.

Security and Limitations

The server runs locally as a subprocess over stdio and has no network listener or hosted remote service. Query and document embeddings use a local model. Storage credentials stay in the environment and are used only for the configured dataset. The default memory:// catalog is ephemeral. Writes are disabled by default and should be enabled deliberately with least-privilege storage credentials. Search table functions are intentionally unavailable through infino_sql; use the dedicated retrieval tools instead.