The Infino MCP server lets a coding agent work with an
Infino Cloud database directly. Once it is connected, you can ask
the agent to create a table from your files, load them, and query them, without
writing any client code yourself.
What you need
- Your endpoint, in the form
https://api.platform.infino.ws/<database>. The last path
segment is the database the agent works in.
- An API key that starts with
inf_. The agent creates tables and writes data, so use a
key with read, write and manage capabilities; a new key from the console has all three by
default. Creating a table needs manage, and adding documents needs write. A read-only key
can search and run SQL reads but is refused every write. For key basics, see
Authentication.
- Node.js 20 or later, because the server runs through
npx.
Claude Code
Export your key, then add the server with one command. Replace <database> with your
database name.
Run claude mcp list to confirm the server is registered.
Other MCP clients
Any client that runs MCP servers over stdio can use the same three pieces: the command
npx, the arguments -y @infino-ai/mcp-server, and two environment variables.
Cursor reads this block from ~/.cursor/mcp.json (all projects) or
<project>/.cursor/mcp.json (one project). For Codex and other clients, put it in your
client’s MCP config. Some clients use a different top-level key than mcpServers, so
check your client’s documentation for the exact shape. Restart the client after you edit
the file.
What the agent can do
With the server connected, the agent can:
- create the database if it does not exist yet (
infino_create_database) and create a
table (infino_create_table), optionally with a vector column for semantic search
- load rows from your files (
infino_add_documents), embedding the text for you
- run SQL (
infino_sql) for counts, filters, joins, and aggregates
- search by keyword (
infino_keyword_search), by meaning (infino_semantic_search), or
both at once (infino_hybrid_search)
Each tool call carries tens of rows, which suits a folder of documents or a small export.
To load a whole corpus, use the CLI or an SDK against the same endpoint. The full
tool list and configuration options are on the MCP server page.
Try it
Paste this into your agent, and replace the path with a folder or file you want to load:
The agent will create the table, load the files, and report what it found.
Next
- MCP server: every tool and setting, including local and bucket storage.
- Authentication: create and rotate API keys.
- Search: how keyword, vector, hybrid, and SQL queries work.
Last modified on November 20, 2026