Make it yours
The loop is an API. Wire every stage into your own systems with the SDKs, the CLI, and an MCP server your coding agents can drive directly.
You've followed one recording through the entire ground-truth loop. The last step is making that loop part of your infrastructure — because a data engine that only works through a UI isn't infrastructure, it's a tool. Avala is built API-first, so every stage you just learned is programmable.
Three ways to drive it
The SDKs. A Python SDK (pip install avala) and a TypeScript SDK (@avala-ai/sdk) cover datasets, projects, tasks, exports, quality, and automation with full type hints. Both talk to a REST API of 50+ endpoints, so anything the platform does, your code can do.
from avala import Client
client = Client()
export = client.exports.create(project="<uid>", format="coco")
The CLI. Manage the loop from your terminal — upload datasets, trigger auto-labeling, create and wait on exports, check rate limits:
pip install avala
avala datasets upload ./run.mcap --dataset <uid>
avala autolabel create --project <uid> --model sam3
avala exports create --project <uid> --format yolo
Webhooks. Event-driven by default. Get notified when consensus is reached, a task is submitted, or an export completes — so the loop can trigger your downstream systems without polling.
Let your coding agent drive it
The newest way to use Avala is to not click at all. Avala ships an MCP server that plugs into Claude, Cursor, VS Code, and other agentic tools, exposing the platform's read operations — and, when you opt in, its writes — as tools an AI agent can call directly.
claude mcp add avala -- npx -y @avala-ai/mcp-server
It's read-only by default; mutations are gated behind an explicit flag, so you can let an agent explore your datasets, check dataset health, and inspect annotations safely, then grant write access deliberately. This makes Avala a headless data engine that fits the way teams are starting to build — describing what they want and letting an agent wire it up.
Where to go next
- Quickstart — get an API key and make your first call: avala.ai/docs/getting-started/quickstart
- SDK reference — the full Python and TypeScript surface.
- MCP setup — connect your coding agent.
- Book a demo — see the loop run on your own data.
The one thing to remember
Plenty of tools will help you look at your data. The hard, compounding work is turning it into ground truth you can trust and closing the loop from a deployed failure back to a better model. That loop is the product — and now you know how it runs, end to end. Go make it yours.
Start building: Read the docs →