PAREL Docs

PAREL is a serverless runtime for AI agents. You describe an agent in one file, agent.yaml, and deploy it. PAREL runs its sessions and keeps their state between messages, calls the model you chose (PAREL-hosted, or a provider on your own key), loads the plugins you listed, and records every turn and step so you can inspect them later.

PAREL is in beta; the API may still change. Agents call models with your own provider keys, billed by the provider; a model you give no key runs on PAREL-hosted models, at the provider's price. PAREL plans meter the runtime by steps (one model call each), always-on connections and hosted model use; the Free plan includes 5,000 steps and $1 of hosted models a month. See pricing.

Quickstart →

Deploy your first agent and chat with it from the terminal.

Philosophy →

Why an agent is a cloud object, and why its spec is one text file.

Concepts →

Agents, instances, sessions, versions, turns and steps.

API Reference →

REST endpoints, auth, execution controls, and rate limits.

CLI Reference →

Every command, with examples.

agent.yaml →

Configuration format and options.

Writing Plugins →

Build custom runtime tools, guards, and capabilities.