Developers¶
Your program keeps the SDK it already uses. You change two settings, the base address and the key, and every call it makes is governed: checked against the key's models, routed, protected by Guardrails, priced and recorded. Every answer carries headers that say what happened, so you never have to guess.
Using a particular SDK, framework or tool?
Client integrations has a page for each one (OpenAI, Anthropic and Google SDKs, LangChain, CrewAI, Cursor, Claude Code, n8n and more), each with the two settings and a minimal call.
There are two jobs here, and they use two different credentials:
| You want to… | You use | Issued on |
|---|---|---|
| Call models from an app, an agent or a script | a key (starts vk_), sent to the gateway |
Daily › Keys |
| Automate Vidai itself: issue keys, onboard teams, read cost | an API token (starts vat_), sent to the Control room's API |
Your account › Personal access tokens, or a service account Enterprise |
A key never manages Vidai, and an API token never calls a model. Keys and API tokens explains which you need.
flowchart LR
app["Your app or agent<br/>(its own SDK)"] -- "key" --> gw["Vidai gateway"]
gw -- "the provider's own format" --> p1["OpenAI, Azure OpenAI"]
gw --> p2["Anthropic, Bedrock"]
gw --> p3["Gemini, Vertex AI"]
gw -. "every call recorded" .-> cr["Control room<br/>Request logs, cost, evidence"]
script["Your automation"] -- "API token" --> cr
Calling models¶
- Find the gateway's address: the address and ready-to-copy SDK settings, in Connect your app.
- Call models through the gateway: curl, Python and TypeScript, with streaming.
- Any SDK to any provider: write to one API, reach a model behind another; what Vidai adapts and how it tells you.
- Model names and aliases: name a model once; let the organisation change what serves it.
- Errors and refusals: every refusal a caller can get, what it means, what to do.
- Request ids and Request logs: quote one id, find the call, see why.
- Fallback and retries: what the gateway retries, which model answered, and how many retries to keep.
- Agents in production: a key per agent and environment, tracing an agent's calls, its own budget and Guardrails.
- Migrate from calling providers directly: the before and after, and what to delete.
- Performance: what the gateway adds to a call's time.
Automating Vidai¶
- Keys and API tokens: keys for apps and agents, API tokens for scripts, service accounts for systems.
- Service accounts and tokens: roles, scopes, rotation and the audit log.
- Event webhooks and scheduled data to your storage: signed events as they happen, and data for BI.
Testing without real providers¶
VidaiMock answers like OpenAI, Anthropic, Gemini and others, streaming and tool calls included, and can fail on purpose (timeouts, rate limits, errors), for free and the same way every time. Point your tests at it directly, or add it to Vidai as a Custom provider (Add a provider) to try routing, fallbacks and budgets without spending on real calls.
Reference¶
The exact contract a program sees: the gateway reference, with endpoints, request headers, response headers, errors, streaming and limits.
For a page per SDK, framework or tool (LangChain, Claude Code, Open WebUI and many more), see Integrations.