lumen-mcp

Drive Lumen's data to SQL to chart to report loop from any MCP client (Claude Code, Claude Desktop, Cursor, VS Code, Goose, ...).

lumen-mcp is a standalone MCP server. It imports Lumen as a dependency and reuses Lumen's own engine; it does not modify Lumen. A couple of not-yet-public Lumen helpers are reached via a small _shims.py, which picks up the public API automatically once the installed Lumen exposes it.

Two modes

• Keyless (default, no API key). The host LLM you are already talking to writes the SQL and the Vega-Lite spec; lumen-mcp runs them through Lumen (DuckDB workspace, spec normalization, rendering, report export). The host is the agent.

• Keyed (opt-in). Lumen's own SQLAgent / VegaLiteAgent / Planner run inside the server. You just describe what you want. Requires an LLM key (see below).

Same tools, same DuckDB workspace, same chart/report output. The key just flips the brain.

Related MCP server: agentic-ops-builder

The session is a DuckDB workspace

Each SQL result is materialized as a real table (via Lumen's DuckDBSource.create_sql_expr_source(materialize=True)), so results accrete in one connection and you reference them by table name. Charts and reports bind to those tables.

Keyless tools

• connect_source(uri, name?) - connect a .db/.duckdb, .csv, .parquet, .json, or :memory:.

• list_tables() / describe_table(table) - schema + a small sample.

• run_sql(sql, name?) - execute; the result becomes table name; returns columns + sample.

• render_vegalite(spec, table) - normalize the spec, render; returns an inline PNG plus saved PNG/HTML paths and a ui_uri.

• refine_chart(chart_id, spec_patch) - deep-merge a patch and re-render under the same id.

• get_chart(chart_id) / list_charts() - fetch or list rendered charts.

• view(target) - show a chart (by id) or a saved .png inline; HTML files return a path to open.

• build_report(items, title, formats?) - assemble charts + markdown into a self-contained HTML and a reproducible .ipynb; returns inline chart previews too.

• save_session(path) / load_session(path) - persist and restore the workspace and its charts.

• launch_dashboard() / stop_dashboard() - serve the session's charts + tables as a live, interactive Lumen dashboard (a background panel serve process) at a localhost URL.

Charts are also served as ui://lumen/chart/{id} MCP-App resources (interactive HTML) for Apps-capable hosts (Claude Desktop/web).

Keyed mode (Lumen's own agents)

Start the server with an LLM key in the environment and one extra tool appears:

OPENAI_API_KEY=...   lumen-mcp     # or ANTHROPIC_API_KEY=...

• lumen_ask(prompt) - Lumen's own Planner + SQLAgent + VegaLiteAgent run headless over the workspace: Lumen writes and runs the SQL and builds the chart itself. Returns the chart inline plus the generated SQL and a summary.

Set LUMEN_MCP_LLM_MODEL to override the default model (gpt-4o / claude-sonnet-4-5).

You can also enable keyed mode at runtime without restarting:

• set_llm_key(api_key, provider, model?) - configure a key mid-session (it passes through the conversation, so prefer the env var for anything sensitive and rotate afterward).

• ui://lumen/setup - an in-chat key-entry pane on Apps-capable hosts (Claude Desktop/web) that submits the key without routing it through the model.

Until a key is configured, lumen_ask returns a clear "not configured" message.

Live dashboard

launch_dashboard() runs a Panel server (inside lumen-mcp, reusing the panel-live-server pattern) that reconstructs the session's charts and tables into a live, interactive dashboard and returns a http://localhost:PORT/... URL. Unlike the static HTML export, its widgets and tables re-query the DuckDB workspace live. stop_dashboard() shuts it down. Requires a local browser (localhost).

Quick start

pip install -e .
python examples/make_sample_db.py          # writes sample.db
# register with your client, e.g.:
#   claude mcp add lumen-mcp -- lumen-mcp

Then, in the client: connect to sample.db, run a GROUP BY query, and render a bar chart.

Development

pip install -e ".[dev]"          # editable install with pytest
pytest                           # run the tests
ruff check src tests examples

Tests: test_slice (keyless logic), test_roundtrip (MCP protocol), test_dashboard (spawns a live server), test_keyed (skips unless an LLM key is set).

Status

Keyless loop + delivery hardening + live dashboard + keyed agentic mode (15 tools). See CHANGELOG.md for details.