Tool Pick | mcp-bi — One Set of Tool Calls to Read Dashboards Across Superset, Metabase, and Power BI

🌏 中文版

Tool Info

What Problem It Solves

Say you're building an operations agent that answers "how does this quarter's revenue compare to last quarter" or "which channel lost the most traffic." The data already lives in dashboards on your company's BI platform — but the agent either has to learn each platform's own query language (SQL for Superset, DAX for Power BI, explores for Looker), or it only gets handed a screenshot of a dashboard. The screenshot path is quietly dangerous: a model looking at an image will fluently "describe" a trend line it never actually measured, with numbers it made up.

mcp-bi collapses that platform difference into one environment variable, BI_BACKEND. The default is a memory mode seeded with three fixture dashboards — no credentials needed, so you can try the tool interface before wiring up anything real. Point it at superset, metabase, powerbi, tableau, looker, qlik, or quicksight instead, and the agent calls the same twelve tools (bi_list_dashboards, bi_chart_data, bi_drill_down, and so on) regardless of which platform sits behind them — swapping vendors doesn't touch the agent's own code. The design rule stated plainly in the README: "numbers are authoritative; pixels are for people." bi_render_chart returns a PNG and the statistics behind it in the same call; bi_insights computes direction, change, extremes, mean, and outliers on its own — the numbers actually worth citing — so narration has something real to point at.

Good fit for: teams already running Superset or Metabase who want an agent to query existing dashboards directly instead of re-wiring a database connection; a support or ops agent that needs to answer "what's the latest number for X"; anyone juggling several commercial BI platforms who wants one tool interface instead of writing a separate integration per vendor.

Quick Start

Install

# Zero-config trial: three seeded fixture dashboards
git clone https://github.com/zavora-ai/mcp-bi
cd mcp-bi
cargo run

# Or install the version published to crates.io
cargo install mcp-bi

Basic Usage

# Point at a real Apache Superset (prefer username/password over a raw token, see below)
BI_BACKEND=superset \
  SUPERSET_URL=http://localhost:8088 \
  SUPERSET_USERNAME=admin SUPERSET_PASSWORD=admin \
  cargo run

Typical call sequence an agent follows:

1. bi_backend_info      # Confirm which platform is connected and what it supports
2. bi_list_dashboards   # List available dashboards
3. bi_get_dashboard     # See a dashboard's drillable dimensions
4. bi_chart_data        # Pull the raw rows behind one chart
5. bi_insights          # Compute direction/change/extremes/outliers to cite in the answer

Advanced Usage

# Drilling down: narrow by filters, break down by a dimension, get row counts
# before/after (invoked by the agent through the bi_drill_down tool, not a CLI flag)

# Spin up a real Superset to verify against, instead of trusting the docs alone
docker run -d --name superset-bi -p 8088:8088 \
  -e SUPERSET_SECRET_KEY=local-only apache/superset:latest
docker exec superset-bi superset db upgrade
docker exec superset-bi superset init
docker exec superset-bi superset load_examples   # loads 9 real example dashboards
python3 scripts/verify-superset.py ./target/release/mcp-bi

Comparison With Existing Approaches

Caveats

• Only Superset has been verified live. The README states Metabase is tested against recorded responses following its documented API, and the five commercial adapters (Power BI, Tableau, Looker, Qlik, QuickSight) have never run against a real tenant — the author's own words: "expect at least one surprise per platform." Verify one of these yourself before relying on it in production.
• License claim is inconsistent. crates.io lists the published crate as Apache-2.0, but GitHub's own license-detection field on the repo shows Other/NOASSERTION, suggesting there's no standard-format LICENSE file in the repository. Check the actual license file before adopting it, not just the badge.
• One day old, one star. The codebase isn't trivial (roughly 3,800 lines of Rust with 27 tests), but community validation is essentially nonexistent — expect to read the source yourself if you hit an issue.

Today's Takeaway

The hard part of BI integration usually isn't reaching the API — it's whether the agent's answer is backed by an actual number. Hand a model a dashboard screenshot and it will fluently narrate a trend it never measured. mcp-bi makes "the chart" and "the statistics behind the chart" a pair that must be returned together, closing off the "narrate from a picture" path at the protocol level.

References

• zavora-ai/mcp-bi GitHub repo: full README — architecture, the 12-tool list, per-platform environment variables, the three findings from live Superset verification, and the "not yet verified" disclosure.
• mcp-bi — crates.io: license (Apache-2.0), version (0.1.1), and line-count stats (3,857 lines Rust / 150 lines Python).
• GitHub API repo metadata (zavora-ai/mcp-bi): stars (1), language (Rust), creation date (2026-09-11), and the license-detection field (Other/NOASSERTION) via the GitHub REST API.