gemini-radar

Gemini API infrastructure FinOps SDK — part of the Hyperscaler Radar suite.

Scan a Gemini Developer API key's project for its model catalog, tuned models, cached content, uploaded files, File Search Stores, and batch jobs. Detect external service relationships from cached-content system instructions (the closest thing this API has to an OpenAI assistant's stored instructions), flag cost/hygiene anomalies via a findings engine, and export inventory to CSV or a draw.io architecture diagram.

Install

pip install gemini-radar

Requires Python 3.10+. CSV and draw.io export ship in the base install.

A different shape than openai-radar / anthropic-radar — on purpose

The Gemini Developer API (generativelanguage.googleapis.com) has no org-wide usage/cost report and no API-key-listing endpoint — a single key just scopes calls to whichever Google Cloud project issued it in AI Studio. So gemini-radar doesn't do token/dollar findings the way its siblings do; instead it watches the resources that carry their own standing cost regardless of billing visibility:

• cachedContents — billed hourly for storage until they expire, whether or not they're being hit
• fileSearchStores — persistent RAG storage
• batches — async job failures that silently waste quota
• files — 48h auto-delete TTL, useful for pipeline hygiene

For actual token spend and Cloud Billing data on the backing project, pair this with gcp-radar rather than duplicating that surface here.

from gemini_radar import RadarClient, Runner

# Reads GEMINI_API_KEY (or GOOGLE_API_KEY) from environment
client = RadarClient()
result = Runner.run_sync(client)

print(result.summary())
result.export_csv("./out/")
result.export_drawio("./out/gemini_arch.drawio")

CLI

# Full scan — findings table to stdout
gemini-radar run

# CSVs + draw.io diagram
gemini-radar run --csv-dir ./out --drawio-file arch.drawio

# JSON instead of a table
gemini-radar run --output json --out-file scan.json

# Skip batch jobs or File Search Stores
gemini-radar run --no-batches --no-file-search

# Findings only
gemini-radar findings

# Print the version
gemini-radar version

Flags follow the Radar suite convention: --output/-o selects table or json, --out-file writes the JSON payload, --csv-dir writes per-resource CSVs.

Findings engine

Service relationship detection

cachedContents is the one Gemini resource that stores a real systemInstruction, so detect_relationships() mines that text (plus tuned-model and file-search-store display names) for the same external-service signal set as the rest of the Radar suite, and emits ServiceRelationship edges (dashed lines in the draw.io diagram):

SDK structure

src/gemini_radar/
├── client.py               # RadarClient (single-key auth)
├── runner.py                 # Runner, RunConfig, RunResult
├── findings.py                # FindingEngine, Finding, Severity
├── relationships.py           # systemInstruction-based service relationship detection
├── models/base.py             # Pydantic v2 models for all resource types
├── scanners/                  # models.py, cached_contents.py, files.py, batches.py
├── exporters/                 # csv_exporter.py, drawio_exporter.py
└── cli.py                     # gemini-radar CLI

Part of the Hyperscaler Radar suite

aws-radar · gcp-radar · azure-radar · oci-radar · openai-radar · anthropic-radar · gemini-radar · datadog-radar · coreweave-radar

License

MIT