Torre — Atlas Control Plane

One screen for a whole MongoDB Atlas fleet, with a Claude assistant grounded in the real clusters — not generic MongoDB trivia. Ask whether an M30 is enough and it answers from your p95 CPU.

Everything comes from the Atlas Admin API v2. UI is in pt-BR; code and docs in English.

The demo in 5 steps

1. Overview — the whole fleet in one snapshot. Clusters, status, cost and alerts, without opening a dozen Atlas tabs.

2. Health Score — one number, and where it came from. 0–100 built from Performance Advisor, COLLSCAN shapes, cluster status and MongoDB version, with the points broken out per component.

3. Scale — the tier answer, from data. 24h CPU (p95/avg), memory, storage and connections, next to the cluster's native auto-scaling status and a tier simulator.

4. FinOps — the bill next to the utilization. Current invoice from the Billing API, estimated cost per cluster, and a verdict per row.

5. AI Chat — grounded in the fleet. Streaming Claude with the cluster context attached, history persisted in Atlas.

Note what it does in that screenshot: asked about 24h, it says it only has the last 5 minutes and shows how to get the rest, instead of making a number up.

Also on the menu: Performance Advisor (suggested indexes, one-click creation via pymongo, Claude analysis, PDF export), Query Profiler (parsed slow queries with a real explain('executionStats')) and Compare (two clusters side by side).

Screenshots run against a live Atlas org; project and cluster names are replaced with neutral ones.

How it fits together

flowchart LR
    UI["React 18 + Vite + LeafyGreen"] -- "/api" --> API["FastAPI (api.py)"]
    API --> AC["atlas_client.py<br/>Admin API v2, TTL caches, tier heuristic"]
    API --> AI["ai_agent.py<br/>Claude chat + PDF"]
    API --> CM["chat_memory.py"]
    AC --> ATLAS["Atlas Admin API v2 + Billing"]
    AI --> CLAUDE["Claude API (streaming)"]
    CM --> DB[("Atlas cluster<br/>chat history")]

Loading

Three deliberate choices:

• Credentials never leave the backend. The frontend only talks to /api.
• The assistant is fenced in. Scope is restricted to Atlas (an "M30" is a tier, never a Kubernetes cluster) and it must separate real API data from pattern-based recommendation.
• Cheap to keep open. Reused HTTP session, TTL caches, and Anthropic prompt caching on the static system block plus a ~2-minute cluster snapshot. Token spend shows at GET /api/metrics.

Run it

Needs Python 3.10+, Node 18+, an Atlas Admin API key and an Anthropic key.

cp .env.example .env    # fill in the keys
./run_react.sh          # API :8765, UI :5290

ATLAS_PUBLIC_KEY=
ATLAS_PRIVATE_KEY=
ATLAS_ORG_ID=
ANTHROPIC_API_KEY=
MONGODB_URI=                  # optional: index creation + chat history
CLAUDE_MODEL=claude-sonnet-5  # optional
API_AUTH_TOKEN=               # optional: protects the API

Override ports with API_PORT=8770 WEB_PORT=5295 ./run_react.sh.

Docker (nginx serves the build and proxies /api):

docker build -t torre . && docker run --env-file .env -p 18085:8080 torre

Tests

python -m unittest discover -s tests -v

24 pure-logic tests — scaling heuristic, injection guards, chat-memory id validation. No credentials needed.

Layout

api.py              FastAPI routes, middleware, auth
atlas_client.py     Admin API v2 client + scaling recommendations
ai_agent.py         Claude analysis, chat, PDF (streaming)
chat_memory.py      Chat history in Atlas
observability.py    Structured logs + /api/metrics
frontend/src/pages/ One component per page

Credits

Based on Maestro by Carime (maestro-atlas-landing-zone).

MIT — see LICENSE.