# How It Works

A cloud control plane for configuration and monitoring, and an agent inside your infrastructure that reads, moves, and loads your records and reports metadata and run results to the control plane.

## Platform · How it works — A cloud control plane. An agent on your infrastructure.

Amfelt is two components. A cloud control plane where you configure, orchestrate, and monitor, and a lightweight agent that runs inside your infrastructure, moving and structuring your data. The two exchange configuration, data contracts, run results, connection credentials, and column statistics for the advisor, and the agent moves your records on your own infrastructure. [Security & Trust](/platform/security/) sets out what crosses.

Links: [Join the waitlist](/waitlist/)

## Two sides, one boundary.

People sign in through the app, machines drive the CLI and API, and a governed Model Context Protocol (MCP) endpoint for AI agents is on the roadmap. All of them work in the control plane: configuring, scheduling, approving, monitoring.

The agent works inside your infrastructure, discovering schemas, relaying data between connectors as Apache Arrow record batches, generating contracts and layers. What crosses between the two includes configuration and contracts, commands, run results, connection credentials, connection-test answers, and column statistics for the table-type advisor, among them the earliest and latest value of date and time columns. [Security & Trust](/platform/security/) sets out each of these.

## The lifecycle — Discover, generate, execute, adapt.

- **Discover**: Point an agent at a source and run discovery. It reads the schemas, tables, columns, and constraints.
- **Generate**: That metadata becomes a data contract (ODCS) for you to commit to Git. The layers are generated from it.
- **Execute**: The agent relays Apache Arrow record batches between connectors into staging, then runs history and bridge in order, as far as the job's layer setting goes.
- **Adapt**: Run discovery again and each schema change against the last discovered contract is graded patch, minor, or major for you to review. Drift policies are on the roadmap.

## Under the hood — The choices that make it predictable and safe.

Four choices built into the platform, rather than into each pipeline.

### Arrow between connectors

Connectors exchange Apache Arrow record batches, which the agent relays unchanged. Each end still reads or writes record by record.

### Deterministic by design

Pipeline SQL renders from templates in your data platform's dialect, so the same contract, configuration, and connector version render the same SQL, apart from each run's ID and start time.

### Capability-based connectors

Every connector declares discover, extract, or load à la carte, and by default runs as a local subprocess on the agent host.

### Context by design

Every sync compares the extracted schema with the source's contract and warns on the differences it finds. With documentation on, a commit from the app regenerates the docs, unless it changes only contracts.

Links: [Platform architecture](/docs/concepts/architecture/)

## See the architecture run.

Join the waitlist to hear when Amfelt opens to more teams.

Links: [Join the waitlist](/waitlist/)