Skip to content

Use Case: AI Assistant User

You have a bundle. You want to ask questions about the data, explore it, and get answers — without writing SQL or Python yourself. Start the MCP server once and your AI assistant (Claude, ChatGPT, or any MCP-compatible assistant) gets direct access to the bundle.

Setup

Run this once in your project:

bundlebase setup-agent

This configures your AI assistant (Claude Code, Cursor, Copilot, or any MCP-compatible assistant) to use Bundlebase automatically — installs the agent skills and wires up the MCP connection. No manual server management.

Use --scope global to configure it once for all your projects instead:

bundlebase setup-agent --scope global

After setup, your assistant has these tools available whenever it works with a bundle:

Tool What it does
status Bundle path, version, row count, last commit
schema Column names and types
sample First N rows as a preview
history Full commit log
query Run any SQL against the bundle

Point your AI assistant at the running MCP server and start asking questions.

Exploring what you have

You don't need to know the schema or what's in the bundle — just ask:

What data is in this bundle? Give me an overview.

Bundle: Q1 2026 Sales — Closed Won
Rows: 18,423
Last updated: v3 — "Added March data" (2026-04-01)

Columns: deal_id, account_name, region, product_sku, amount (Float64),
         close_date (Date), sales_rep, customer_tier

Data spans January through March 2026. All records are closed-won deals.

What changed between the first and latest version?

v1 (2026-01-15): Initial Q1 export — January data only (6,102 rows)
v2 (2026-02-12): Added February data (11,847 rows)
v3 (2026-04-01): Added March data (18,423 rows)

Running analysis

Which region had the highest average deal size in Q1?

EMEA led with an average deal size of $84,200, followed by North America
at $71,400 and APAC at $58,900.

Top 3 regions by average deal size:
  EMEA      $84,200   (3,102 deals)
  NA        $71,400   (8,841 deals)
  APAC      $58,900   (4,210 deals)

Show me the top 10 deals by amount, with the account name and close date.

| deal_id | account_name        | amount     | close_date  |
|---------|---------------------|------------|-------------|
| D-9921  | Acme Corp           | $480,000   | 2026-03-28  |
| D-8847  | Globex Industries   | $412,500   | 2026-02-14  |
| ...     | ...                 | ...        | ...         |

Is there any seasonality in deal close dates — do more deals close at end of month?

Yes, noticeably. 34% of deals close in the last 3 days of each month,
versus an expected ~10% if evenly distributed. This pattern holds across
all three months and all regions.

Digging deeper

Are there any data quality issues I should know about?

A few things worth flagging:

1. 47 records have amount = 0.00 — likely test records or data entry errors
2. customer_tier is null for 312 records (1.7% of total)
3. 3 records have close_date in December 2025, which is outside the Q1 2026 range

Filter out the zero-amount records and the December deals. How does that change the total pipeline value?

After filtering:
  Records removed: 50
  Remaining rows: 18,373
  Total pipeline value: $1,312,847,200 (was $1,312,847,200 — zero-amount records
  didn't affect the sum, but the December deals removed $48,200)

Updating the bundle

The assistant can also make changes when given the right permissions:

Fetch the latest data from the source and commit it as "April refresh".

Fetched 4,218 new records from the configured source.
Bundle updated: 22,641 total rows.
Committed as v4: "April refresh"

Add an index on region so regional queries are faster.

Index created on 'region'. Queries filtering by region will now use the
index instead of scanning the full dataset.

Generating reports

Ask the assistant to produce a formatted PDF report from the bundle — it writes the report template and runs bundlebase generate-report for you:

Generate a PDF report on Q1 sales performance. Include a bar chart of deals by region,
a trend line of deal size over the quarter, and a table of the top 20 deals by amount.
Save it as q1-sales-report.pdf.

I'll create a report from the sales-q1 bundle. Here's what I'm generating:

  - Bar chart: deal count by region
  - Line chart: average deal size by week across Q1
  - Table: top 20 deals by amount with account name and close date

Writing report template and running bundlebase generate-report...

Done. q1-sales-report.pdf saved — 3 charts, 18,423 rows of live data,
generated at 2026-04-05 14:32.

You can be more specific about what you want:

Add a section breaking down deals by customer tier, and filter the trend chart
to North America only. Regenerate the report.

Updated report generated. Changes applied:
  - Added "Deals by Customer Tier" horizontal bar chart after the region chart
  - Trend line now filters to region = 'NA' (8,841 deals)

q1-sales-report.pdf overwritten.

Reports pull live data from the bundle at generation time — if the bundle is updated next month, regenerating the report picks up the new data automatically.

Session recovery

If a previous session failed mid-run, the bundle is always in the last committed state — no partial data to clean up. Just ask:

What's the current state of the bundle? Did the last update complete?

Bundle is at v3 (2026-04-01): "Added March data" — 18,423 rows.
This is a clean committed state. If a previous session failed, it left no
partial data behind.

When to use MCP vs. the Python API

Use MCP (this use case) when:

  • You want to explore, query, or update a bundle through conversation
  • You don't want to write code
  • Your AI assistant supports MCP tools natively

Use the Python API when:

  • You're building automation or pipelines that run on a schedule
  • You need programmatic control over exactly what runs
  • You want to integrate bundle operations into existing code

Next steps