start

Frequently asked questions

Get direct answers about choosing a Kestrel product, durable work, model access, deployment, and when Kestrel is unnecessary.

Suitebeginner0.7.0 Stable
Verified 2026-07-20View sourceReport a docs issue

Which Kestrel product should I start with?

Use Desktop for work in local folders and repositories. Use Kestrel One when an invited team needs shared Threads, Projects, and Knowledge. Use the SDK or HTTP interfaces when Kestrel should run behind an application you already own. The journey chooser compares the first steps.

Do I need to run a database for Desktop?

No. Kestrel Desktop includes Local Core and manages its own persistence. You choose a model provider and a project folder during first run.

Can anyone sign up for Kestrel One?

Kestrel One is currently available to invited teams. Use the invitation from your organization administrator to sign in or create the invited-email account, then explicitly join the organization. Public signup and self-hosted consumer onboarding are not currently documented.

Why not call a model directly?

A direct call is simpler for a stateless answer. Kestrel becomes useful when work must continue across requests, wait for a person, produce artifacts, support operator intervention, or leave enough evidence to explain and replay later.

What is the runner service?

It is the network-hosted form of the Kestrel runtime. A trusted application server calls it with authenticated user and session context. It exposes the same Execution Protocol that Local Core provides locally, but through HTTP and streaming endpoints. A browser should not receive its bearer token or provider credentials.

For local development, kestrel web provides a compatible HTTP URL by bridging requests to Local Core. It does not start a second runtime.

What is the difference between streaming and subscriptions?

Streaming follows the request you just started. A subscription follows selected background events for a session, Thread, or run. Replay examines persisted evidence after execution ends.

When should I use OpenAI-compatible HTTP instead of the SDK?

Use compatible HTTP when an existing server already speaks OpenAI-style chat or responses APIs. Use the SDK when you want typed agents, sessions, memory, streaming, waiting, resume, cancellation, and richer Kestrel controls.

When is Kestrel too much?

Kestrel may be unnecessary when a feature only sends a prompt and shows an answer, with no durable state, recovery, operator workflow, or later audit requirement.

How do I deploy it?

Run the runner service behind your application server, configure authenticated server-to-server access, then verify the complete request and recovery path. For local development, expose Local Core with kestrel web. Start with Operate for the distinction between local and remote environments.