MidSphere

Autonomous Context Circuit-Breaker Engine for Data Platforms
DataHub · MCP · Google Antigravity SDK · Gemini

Organizations deploy AI agents faster than they can keep catalog context trustworthy.

Schema drift, quality defects, ungoverned PII, and missing ownership do not stop at an alert. Agents that continue to read the catalog inherit the damage.

Many catalogs only record that something failed. MidSphere treats that as insufficient. It runs a closed loop that audits context, scores selective blast radius, applies advisory quarantine, produces a dry-run fix artifact, and leaves an MBOM attestation path so the next human or agent can inherit safer knowledge.

When DataHub mutations are enabled and GMS is reachable, quarantine is written as graph state (tags, structured properties, documents). Otherwise it remains job-local with mutations_applied=false. MidSphere Consumer Gate returns BLOCKED for MidSphere-aware clients. This repository does not claim that every third-party MCP consumer is automatically denied by the platform.

Built for Build with DataHub: The Agent Hackathon.

Why not a scanner or a gateway

MCP security scanners and gateways inspect tool metadata and block at the proxy. MidSphere starts where those stop.

Capability	Scanners and gateways	MidSphere
Detection surface	Tool metadata and tool calls	Catalog context: quality, PII, freshness, ownership
Action on finding	Alert or proxy deny	Advisory quarantine written to DataHub graph state (tags, structured properties, documents)
Reversibility	Reconfigure the proxy	Lift via remove_tags through the official MCP server
Consumer enforcement	At the gateway	Consumer Gate returns BLOCKED for MidSphere-aware clients only (no platform-wide claim)
Evidence	Scan report	MBOM attestation and dry-run fix artifact

Live surfaces

Surface	URL
Command Center	https://midsphere.vercel.app
API (Cloud Run)	https://midsphere-api-1053689941073.us-central1.run.app
Source	https://github.com/adi-IL/MidSphere

Status honesty (Command Center)

The UI is minimal monochrome (no emoji). Status labels are defined as follows.

UI label	Meaning
API · Job	API is up and the canvas job came from the API (latest, Run, or Trip)
API · Sample	API is up but the canvas is a local demo fixture (e.g. scenario tab without Run)
Offline	API unreachable; sample fixtures from frontend/src/lib/demoJob.ts
GMS / GMS off	API GMS probe (datahub_reachable)
Pipeline Sample / Agent / Live MCP	Run kind derived from execution_mode and job source
Run completes	Closed loop finished (Google Antigravity primary and/or deterministic fallback)
Decision Fix / Attestation	Scenario paths under examples/ (dry-run)
Live tags / save_document / Lift remove_tags	GMS reachable and TOOLS_IS_MUTATION_ENABLED=true (see job mutations_applied and TOOL events)

Operator path: open the live app, select Hospital, Retail, or Taxi, click Run, inspect the selective graph (impacted vs open fork), use Decision Findings / Gate / Attestation, then Trip / Lift. Expand Telemetry for stage logs. When mutations are live, look for add_tags official_mcp ok after trip/Run and remove_tags official_mcp ok after Lift (also in the Network response body).

Manual smoke checklist:docs/MANUAL_TEST.md.

Closed loop

Stage	Component	Shipping behavior
1. Audit	ContextSentinel	Scenario findings (quality, PII, freshness)
2. Topology	CascadeAnalyzer	Selective blast radius and risk score
3. Break	CircuitBreaker	Advisory quarantine; live tags only if GMS and mutations are enabled
4. Fix	FixArchitect	Dry-run SQL/dbt path under examples/prs/
5. Inherit	MBOMEngine	MBOM path under examples/mboms/; live save_document only if mutations are enabled

flowchart LR
T([Trigger]) --> S1["ContextSentinel<br/>1 · Audit"]
S1 --> S2["CascadeAnalyzer<br/>2 · Topology"]
S2 --> S3["CircuitBreaker<br/>3 · Break"]
S3 --> S4["FixArchitect<br/>4 · Fix"]
S4 --> S5["MBOMEngine<br/>5 · Inherit"]
MCP[["DataHub MCP<br/>when GMS is configured"]]
MCP -. "read: list_schema_fields" .-> S1
MCP -. "read: get_lineage" .-> S2
MCP -. "write: add_tags / remove_tags" .-> S3
MCP -. "write: save_document" .-> S5
classDef stage fill:#ffffff,stroke:#333333,stroke-width:1px,color:#111111;
classDef mcp fill:#f5f5f5,stroke:#333333,stroke-width:1px,stroke-dasharray:4 3,color:#111111;
class T,S1,S2,S3,S4,S5 stage;
class MCP mcp;

Loading

Primary workforce:Google Antigravity SDK for Python (pip install google-antigravity; Agent + LocalAgentConfig with custom MidSphere stage tools on Vertex Gemini). Official DataHub MCP attaches when GMS is reachable. Stage tools may call get_lineage, list_schema_fields, add_tags, and save_document when mutations are enabled. On Agent failure, a deterministic fallback completes the job with simulated telemetry labels when offline.

Scenario packs

Packs are defined in backend/agents/scenarios.py and mirrored in Command Center fixtures.

Scenario	Root dataset	Risk	Quarantine target	Open fork
Hospital (healthcare)	healthcare.raw_patients	61	mart_billing	mart_demographics
Retail (ecommerce PII)	ecommerce.raw_customers	88	mart_customer_360	mart_product_catalog
Taxi (NYC freshness)	nyctaxi.raw_yellow_trips	65	mart_monthly_revenue	mart_trip_distance_stats

Timeline (graph control, off by default) is a closed-loop narrative scrubber. It is not a live GMS history browser.

Stack

Layer	Technology
Frontend	React 18, Vite, TypeScript, Tailwind, React Flow, Geist
Backend	FastAPI, Pydantic, httpx, uvicorn
Primary workforce	Google Antigravity SDK (Python package google-antigravity) + backend/agents/stage_tools.py
Fallback	Deterministic stages in backend/agents/pipeline.py
Models	Gemini via Vertex AI (ADC)
Catalog	DataHub GMS and official mcp-server-datahub via uvx
Deploy	Vercel project midsphere, GCP Cloud Run midsphere-api

MCP tools (real names only)

Read:search, get_entities, list_schema_fields, get_lineage, get_lineage_paths_between, get_dataset_queries

Write (requires TOOLS_IS_MUTATION_ENABLED=true on the MCP server): add_tags, remove_tags, add_structured_properties, save_document

No invented datahub_* tool prefixes. No MCP create_incident.

Quick start

git clone https://github.com/adi-IL/MidSphere.git
cd MidSphere
cp .env.example .env
# optional: DATAHUB_GMS_URL / DATAHUB_GMS_TOKEN# Vertex: GOOGLE_CLOUD_PROJECT and ADCcd backend
python3 -m venv .venv &&source .venv/bin/activate
pip install -r requirements.txt
export ANTIGRAVITY_ENABLED=true
export USE_OFFICIAL_MCP=true
PYTHONPATH=. uvicorn app.main:app --reload --port 8000
# second terminalcd frontend && npm install && npm run dev

Cloud Run cannot use laptop localhost as GMS. Operators run official DataHub quickstart on GCE and set DATAHUB_GMS_URL to the VM public IP (see scripts/README.md).

./scripts/gce_datahub_schedule.sh apply # weekday start / nightly stop
./scripts/gce_datahub_schedule.sh stop-now
./scripts/gce_datahub_schedule.sh start-now # starts VM, re-points Cloud Run, waits for GMS

API

Method	Path	Purpose
GET	/v1/health	Health and stack metadata
POST	/v1/jobs	Run closed loop
GET	/v1/jobs/latest	Latest job (404 if none; does not auto-run)
GET	/v1/jobs/{id}	Job by id
POST	/v1/jobs/{id}/lift	Lift quarantine
POST	/v1/jobs/{id}/trip	Re-run closed loop with auto-trip
GET	/v1/gate/{urn}	Consumer Gate decision
GET	/v1/gate-default-blocked	Convenience gate for healthcare mart_billing only (not “latest job blocked entity”)

Sample Datasets

To ensure MidSphere's agents operate on realistic, complex lineage graphs, we developed three custom data environments for this project.

Because these datasets contain raw sample data and ingestion recipes that would bloat the core application repository, they are hosted separately.

Accessing the Datasets

You can view and download the full dataset schemas, mock data, and DataHub ingestion recipes here: datahub-project/static-assets/tree/main/datasets

The Environments

1. Healthcare (healthcare/):

   • Simulates a hospital billing and demographics pipeline.
   • Planted Issue: Negative billing amounts.
   • MidSphere Action: Quarantines mart_billing, leaves mart_demographics open.

2. Retail (fiction-retail/):

   • Simulates an e-commerce platform with customer 360 and product catalog marts.
   • Planted Issue: Unmasked SSN (PII leak).
   • MidSphere Action: Quarantines mart_customer_360.

3. NYC Taxi (nyc-taxi/):

   • Simulates a high-volume transit data pipeline.
   • Planted Issue: Data freshness / delayed ingestion.
   • MidSphere Action: Quarantines mart_monthly_revenue.

If you are running the GCE VM Quickstart, you can use scripts/load_healthcare.sh in conjunction with these datasets to populate your live DataHub instance.

Open-source contribution

Upstream skills PR (open):
datahub-project/datahub-skills#88 (datahub-circuit-breaker, datahub-quality-circuit).

Local mirrors: skills/.

Repository layout

Public OSS tree aligned with common monorepo practice and the hackathon submission requirements (public source, Apache 2.0, setup instructions, optional sample artifacts):

.
├── LICENSE Apache 2.0 (required; visible at repo root)
├── README.md Overview, honesty table, quick start
├── ARCHITECTURE.md Shipping contracts (short)
├── .env.example Env template (no secrets)
├── .gitignore
├── .github/ CI + deploy workflows
│ ├── DEPLOY.md
│ └── workflows/
├── backend/ Closed-loop API (Cloud Run)
│ ├── app/ api/ agents/ gate/ midsphere_mcp/ models/ tests/
│ ├── requirements.txt
│ └── Dockerfile
├── frontend/ Command Center (Vercel)
│ ├── src/ public/
│ └── package.json
├── examples/ Judge-readable sample outputs (optional but present)
│ └── mboms/ prs/ jobs/ telemetry/
├── skills/ Upstream skill mirrors (datahub-skills PR #88)
├── scripts/ Operator helpers (GCE GMS schedule, bootstrap)
└── docs/ Index + MANUAL_TEST.md (judge smoke)

Path	Role for judges
Live Command Center	End-to-end UI
examples/	Offline MBOMs, dry-run diffs, telemetry without running the stack
backend/ + frontend/	Full source to run locally
LICENSE	Apache 2.0

Local-only material (not part of the public product surface) is gitignored: resource/, internal/, large recon dumps.

Documentation

1. ARCHITECTURE.md: shipping system contracts
2. docs/MANUAL_TEST.md: judge / operator manual smoke
3. examples/README.md: scenario artifacts
4. backend/README.md, frontend/README.md
5. .github/DEPLOY.md: deploy isolation

License

Apache License 2.0. See LICENSE.