DataHub Steward Squad
01 / 08
Build with DataHub · The Agent Hackathon

DataHub Steward Squad

A multi‑agent team that turns DataHub context into governed action.
◆Real MCP loop ◆Claude‑powered ◆Approval‑gated ◆Apache‑2.0
The gap

DataHub already knows everything

⛓

Owners & lineage

Who owns what, and what it feeds.

✓

Quality & assertions

Freshness and SQL checks, pass/fail.

◈

PII & glossary

Sensitive fields and business terms.

What's missing is an agent that turns that context into safe, governed change.
The team

One Chief Steward, five specialists

◎

Catalog Scout

Finds assets & coverage gaps.

⛓

Lineage Investigator

Traces blast radius.

⚑

Quality Sentinel

Failing assertions & PII gaps.

✍

Stewardship Writer

Governed writeback proposals.

⬢

Release Captain

Packages evidence & SQL.

★

Chief Steward · Claude

Reasons into the brief.

How it works

A real loop over the Model Context Protocol

01
Read
search · get_entities · get_lineage
→
02
Analyze
5 agents · grounded findings
→
03
Write back
update_description · add_tags
→
04
Verify
re‑read & diff
Every step runs against real DataHub via the official mcp‑server‑datahub — approval‑gated, never auto‑applied.
Live · on real DataHub

One command, real governance

0
assets read
0
findings surfaced
0
critical assertion
0
writebacks verified
missing ownersfailing assertions unclassified PIIlineage blast radius
It writes back — and proves it

Not a plan on disk. A real mutation.

mcp-demo --live --apply
Applied writeback via MCP mutation tools:
  MCP-002 add_tags: applied

Verification (re-read through MCP):
  finance.fct_revenue · customer_email.tags
     []  →  ['PII']
The PII tag now lives on a real Snowflake column in DataHub — confirmed by re‑reading through MCP.
Genuinely agentic

Claude writes the executive brief

“fct_revenue is failing a critical quality check and feeding certified downstream assets — remediate and tag PII before anything ships.”
It reasons only over findings the agents detected — it invents nothing. No API key? A deterministic engine writes the same brief, so the demo never breaks.
Why it fits

Real use of DataHub. Real originality.

✓Official mcp‑server‑datahub — not a reimplementation
✓Approval‑gated governance, verified writebacks
✓Zero‑credential offline fallback for judges
✓Apache‑2.0 · 29 tests · CI green
❯github.com/maharshi-coding/datahub-steward-squad
Space / → to advance