MCP setup
nlqdb is the memory MCP server: a real database your agent can
GROUP BY / JOIN / aggregate over in natural language, not just a recall
store. It speaks MCP (Model Context Protocol) so any host that supports it —
Claude Desktop, Cursor, Zed, Windsurf, VS Code Continue, Cline — can call
nlqdb_query, nlqdb_list_databases, nlqdb_describe, nlqdb_remember
(write a typed memory row your agent can later query) and
nlqdb_connect_database (attach an existing ClickHouse/Postgres) directly.
Install paths
Section titled “Install paths”Three paths, all terminating at the same /v1/ask orchestration and the
same tools:
- Hosted connector URL (default) — point your host’s MCP config at
https://mcp.nlqdb.com/mcp(the protocol lives at/mcp, not the bare domain); OAuth opens in your browser on first tool call. Zero local setup. - One-click buttons on nlqdb.com (also on
/integrations) — deep-link installs
for Cursor and VS Code, a copy-paste
claude mcp addcommand for Claude Code, and paste-ready config blocks for Claude, Windsurf, Zed, and Codex. npx -y @nlqdb/mcp+ anNLQDB_API_KEYenv var — the local stdio server, authenticated by a key instead of a browser. The headless path: CI, Docker, air-gapped, and coding agents that can’t click a consent screen.
Add it to your host
Section titled “Add it to your host”Claude Code, from a terminal:
claude mcp add --transport http nlqdb https://mcp.nlqdb.com/mcpHosts that take a JSON config block (Claude Desktop custom connectors, Windsurf, Zed — exact key names vary by host; the /integrations page renders the verified per-host block to copy):
{ "url": "https://mcp.nlqdb.com/mcp" }Headless (no browser) — create an sk_mcp_… MCP key at
app.nlqdb.com/app/keys (scoped to MCP, bound to
one host + device, revocable on its own), then:
{ "mcpServers": { "nlqdb": { "command": "npx", "args": ["-y", "@nlqdb/mcp"], "env": { "NLQDB_API_KEY": "sk_mcp_…" } } }}The CLI can list which supported hosts are installed on your machine with
nlq mcp detect. A one-command config writer (nlq mcp install) is not
shipped yet — it lands together with CLI sign-in (device flow); today the
command prints a deferral hint pointing at the dashboard.
Usage from inside the host LLM
Section titled “Usage from inside the host LLM”The agent never sees “create a database”; the DB materialises on first reference:
User: "Remember I prefer metric and I'm vegetarian."Claude → nlqdb_query("preferences", "remember: metric units, vegetarian") → { ok, db: "preferences-93b" }
[next session]User: "Plan me a Berlin food trip."Claude → nlqdb_query("preferences", "what do you remember about me?") → "metric units, vegetarian"