querywise-mcp

An MCP server (and a CLI) that lets an LLM query your databases in natural language through a business semantic layer — glossary, metric definitions, data dictionary, knowledge base, and example queries — grounded against your real schema.

It's a refactor of QueryWise (a full-stack text-to-SQL app) into a headless tool: no web UI, no Postgres requirement. The metadata store is an embedded SQLite + sqlite-vec database, so the server runs from a single file.

Two ways to use it

1. As an MCP server — Claude (or any MCP client) calls the tools. The recommended loop is: get_semantic_context(connection, question) → the model writes a read-only SELECT → run_sql(connection, sql). The client's own model does the reasoning; the server provides grounded context + safe execution.
2. As a CLI — querywise ask <connection> "<question>" runs the full server-side NL→SQL pipeline (compose → validate → execute → interpret). This path needs an LLM provider key (or local Ollama).

The semantic layer, connectors, and execution are shared by both.

Install

python3 -m venv .venv && source .venv/bin/activate
pip install -e .                 # core (SQLite store, sqlite-vec, Postgres + SQLite targets)
pip install -e ".[llm]"          # + Anthropic/OpenAI for `ask` and cloud embeddings
pip install -e ".[bigquery,databricks]"   # + extra target connectors

Configuration is via environment variables / .env (see .env.example). Zero config works for keyword-only operation; add a key (or Ollama) to unlock embeddings and the ask pipeline.

Quick start (zero external infra)

querywise init                                   # create ~/.querywise/querywise.db
querywise connections add shop \
    --connector-type sqlite -c /path/to/app.db   # introspects + embeds
querywise context shop "revenue by segment"      # see the grounded context
querywise sql shop "SELECT ..."                  # run read-only SQL
querywise ask shop "what is total revenue by segment?"   # full pipeline (needs LLM)

Run as an MCP server

querywise serve            # stdio (for Claude Desktop / Claude Code / Cursor)
querywise serve --http     # Streamable HTTP on MCP_HOST:MCP_PORT (default 127.0.0.1:8077)

Register with Claude

First make sure the store the server will read is initialized (and optionally seeded):

querywise init                          # create ~/.querywise/querywise.db
querywise seed-sample                   # optional: zero-infra IFRS-9 sample → connection "ifrs-db"

Use an absolute command path. MCP clients launch the server with a minimal PATH, so the bare querywise-mcp often won't resolve. Point at the entry point inside your venv, e.g. /path/to/.venv/bin/querywise-mcp.

The server won't read your repo .env. It runs from the client's working directory, so pass everything it needs (DATABASE_URL, provider keys, model) in the env block below.

Claude Desktop — edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), then fully quit and reopen Claude Desktop:

{
  "mcpServers": {
    "querywise": {
      "command": "/path/to/.venv/bin/querywise-mcp",
      "env": {
        "DEFAULT_LLM_PROVIDER": "ollama",
        "DATABASE_URL": "sqlite+aiosqlite:////Users/me/.querywise/querywise.db"
      }
    }
  }
}

Claude Code — one command:

claude mcp add querywise /path/to/.venv/bin/querywise-mcp \
  -e DEFAULT_LLM_PROVIDER=ollama \
  -e DATABASE_URL=sqlite+aiosqlite:////Users/me/.querywise/querywise.db
# verify: claude mcp list   (or /mcp inside a session)

Note the four slashes in the SQLite URL — sqlite+aiosqlite:// (scheme) plus the absolute path /Users/me/....

Why DEFAULT_LLM_PROVIDER? It's a server setting, not your chat model. Claude is the client LLM — it calls the granular tools and writes the answer, so it needs no provider config. The server only uses a provider for two things: embeddings (semantic search over your metadata — optional; degrades to keyword-only without one) and the all-in-one ask/generate_sql tools (which run their own LLM). Set it to ollama for key-free local embeddings, or to anthropic/openai (with the matching *_API_KEY in env) if you want to call the server-side ask tool. Omit it entirely to run keyword-only.

MCP surface

Tools (25): list_connections, create_connection, test_connection, introspect_connection, delete_connection, list_tables, describe_table, get_semantic_context, run_sql, generate_sql, ask, query_history, glossary/metric/dictionary/sample-query/knowledge management (list_*/add_*/delete_*, plus add_knowledge_url).

Query paths — the four tools people mix up:

Tool(s)	LLM key?	What it does
get_semantic_context + run_sql	No	Server grounds the question; the client writes the SELECT; run it read-only.
generate_sql	Yes	Server writes SQL from the question but does not execute — review, then run_sql.
ask	Yes	Full pipeline: ground → generate → execute → interpret, returns a Markdown answer.

Resource: querywise://{connection}/schema — the cached schema as text. Prompt: text_to_sql(connection, question) — scaffolds the ground→write→run loop.

connection accepts a connection name or id everywhere.

Connectors

Target	Notes
SQLite	Read-only (mode=ro), zero infra. Great for local files + demos.
PostgreSQL	asyncpg, read-only transaction.
BigQuery	optional extra; service-account JSON in the connection string.
Databricks	optional extra; Unity Catalog or Hive metastore.

All execution is read-only: a static SQL blocklist (DDL/DML/admin/injection) plus connector-level read-only enforcement.

How the semantic layer works

For each question the context builder selects minimal relevant context via a hybrid of (1) vector similarity over embeddings, (2) keyword matching, and (3) foreign-key expansion, then resolves glossary terms, metrics, dictionary value-mappings, knowledge excerpts, and example queries into a structured prompt block. Embeddings are stored as float32 BLOBs and searched with sqlite-vec's vec_distance_cosine; if the extension can't load, search transparently falls back to in-process cosine. With no embedding provider, it degrades to keyword-only matching.

Building the semantic layer

The glossary, metrics, value dictionaries, sample queries, and knowledge docs are populated through the MCP management tools — so you can build them conversationally from an MCP client like Claude, no CLI required. Asking Claude to "add a glossary term active customer defined as … with SQL …" calls add_glossary_term; the same goes for add_metric, add_dictionary_entry, add_sample_query, and add_knowledge / add_knowledge_url (and the matching list_* / delete_* tools to review or remove them). For a ready-made example, querywise seed-sample loads the bundled IFRS 9 banking layer.

Architecture

MCP client (Claude/…)  ──stdio/http──┐
CLI (`querywise ask`)  ──in-process──┤
                                     ▼
                          server.py / cli.py
                                     │
        ┌────────────────┬──────────┴───────────┬──────────────┐
        ▼                ▼                      ▼              ▼
   semantic/        services/               llm/          connectors/
 context builder   query pipeline      agents+providers  PG/SQLite/BQ/DBX
        │                │                      │              │
        └──────── db/ (SQLite + sqlite-vec metadata store) ────┘

Development

ruff check src/
python -m compileall src/

The metadata schema is created on startup (db/init.py) — no migration tool. Switching embedding providers/dimensions clears now-incompatible vectors automatically.