docs
/
Vibe Studio

Vibe Agent

An AI coding agent that builds against your own data, with multi-provider models, streaming and usage controls.

Vibe Agent is the AI coding agent. You describe what you want; it builds it against your actual data model rather than a generic scaffold, because it runs inside the same platform that holds your records.

Multiple model providers

Vibe Agent is not tied to one vendor. Anthropic Claude, OpenAI and DeepSeek are all supported, and which one serves a request is configuration rather than a rewrite.

That matters for three practical reasons: you can pick the model that suits a task, you are not exposed to a single provider's outage, and you can move as pricing and capability change.

StreamingResponses stream as they are generated, so long answers start appearing immediately.
ReasoningModels with extended thinking are supported, including their reasoning output.
VisionMultimodal input — send images alongside text for vision-capable models.

Prompt caching

For providers that support it, prompts are cached. Repeated context — a long system prompt, a schema, a document you keep referring to — is not re-billed on every turn. On agent workloads, which resend a lot of stable context, this is the difference between viable and expensive.

Usage, rate limits and plans

Because the agent is metered per organization, the platform enforces limits rather than leaving you to discover them on an invoice:

ControlWhat it does
Usage trackingRecords tokens consumed per organization
Rate limitingCaps request rate so one workload cannot starve others
Plan policyTies available models and limits to the organization's plan
Provider poolingSpreads work across configured provider credentials

Drop-in API compatibility

Vibe Agent exposes an endpoint that is compatible with the Anthropic Messages API. An existing client built against the Anthropic SDK can be pointed at it and work unchanged.

The difference is what happens around the call: it is brokered by the platform, so tenancy, usage metering, rate limits and plan policy all apply, and the model credentials stay server-side rather than being shipped to a client.

Why this matters

You keep your existing code and tooling, and gain per-tenant metering and control you would otherwise have to build.

Streaming

Streaming is two steps rather than one long request: you open a stream and receive an id, then connect to that stream to receive tokens. The split means a client that loses its connection can reconnect to a stream already in progress instead of starting over.