Getting started

Core concepts

Five nouns carry the whole product. Everything else is a property of one of them.

Connections#

A credentialed link to an external system: a model provider, a database, an HTTP API, a SaaS platform, an MCP server, a vector store or object storage. Secrets are encrypted with AES-256-GCM the moment you save them and are never displayed again. Connections carry network policy too — private hosts are flagged for IP allow-listing or a private runner.

Every connection knows its blast radius: how many workflows and tools stop working if it goes down. That number is on the connection, not buried in a dependency graph, because it is what you need during an incident.

Agents#

A reusable AI worker definition: which model connection it uses, its system prompt, temperature, iteration ceiling, and which tools it may call. Policies attach per agent — memory scope, retry with backoff, step and run timeouts, and whether it requires approval.

Tools#

A typed capability an agent can invoke. Every tool declares an input and output schema, a timeout, a retry policy, idempotency semantics and a risk level from low (read-only) to critical (irreversible — payments, sends, deletions). Risk is not decoration: policy can require an approval gate for anything above a threshold, per environment.

Approvals#

A node that pauses a run for a human decision. The run checkpoints and parks — no compute is consumed while it waits. Both approve and reject are recorded with actor, timestamp and the context the decision was made on.

Workflows and runs#

A workflow is the versioned graph. Starting it creates a run — the durable, event-sourced execution you can watch live, scrub through afterwards, and resume from any checkpoint.