How do I give developers access to multiple AI models?
Summary
- Stop handing out a separate API key for each provider. Put every model behind one governed endpoint and give developers one credential, one API and one place to see what they can use.
- Pick an API that hides provider differences. Unity Gateway's unified APIs are compatible with OpenResponses and OpenAI Chat Completions and translate each request to the model's native format, so switching models doesn't mean rewriting client code.
- Govern models the way you govern data. On Databricks, model services and external provider services are Unity Catalog securables, and a developer needs
EXECUTEto call one. - Offer tiers on purpose: a standard model service for everyone and a separate one for power users, each with its own rate limits.
- For coding agents, the Unity Gateway CLI (
ug) connects supported agents to the same governed models, and developers switch models mid-session with/model.
How do I give developers access to multiple AI models?
Route all model traffic through one gateway, register each provider there once, and give developers a single credential and a provider-independent API. Then decide who can call which models with permissions, not with key distribution. Developers get to pick the right model for each task, and the platform team keeps one place for access, limits, usage and spend.
A key per provider per team breaks down fast: keys get copied into notebooks and pipelines, and nobody can say which team is spending what.
What multi-model access needs
| Requirement | What to look for |
|---|---|
| Model coverage | Frontier proprietary models and open-weight models, hosted for you or reached through your own provider accounts |
| One API surface | A provider-independent API, with native provider APIs available when a team needs provider-specific features |
| Central credentials | Provider keys stored once, encrypted, never handed to developers |
| Per-model permissions | Grant and revoke access per user, group or service principal |
| Capacity controls | Request and token rate limits per service, per user and per group |
| Usage visibility | Token counts, latency and requester for every call, queryable with SQL |
| Coding agent support | Agents reach the same governed models, not separate accounts |
How Databricks approaches multi-model access
Databricks puts all of this in Unity Gateway, which governs model inference, tools and skills from one control plane on Unity Catalog.
A broad model catalog in system.ai. Databricks hosts frontier proprietary and open-weight models and exposes them as ready-to-use, pay-per-token system-provided model services in the system.ai schema. Models you already pay for elsewhere connect through a model provider service: you register the provider once, Databricks encrypts and stores its credentials, and many users and model services can route to it without seeing the key. A Custom provider type takes an endpoint base URL plus a bearer token or API key header.
One API, many models. The unified APIs include the Unified Responses API, which is OpenResponses-compatible and is the one Databricks recommends, and the MLflow Chat Completions API, which works with the OpenAI Python SDK. Databricks translates each request to the downstream model's native format. When a team needs provider-specific features, provider-native APIs are available, and ai_query runs batch inference from SQL or Python.
Access by permission. To query a model service, a user needs EXECUTE on it plus USE CATALOG and USE SCHEMA on its parent. Model services run with the owner's privileges, so callers don't need direct access to the underlying destinations. To give a wide user base and a smaller group of power users different models and quotas, the docs describe a separate model service for each group, each with its own grants.
Stable names for applications. A custom model service gives an application a model-agnostic endpoint, so you change the underlying model without changing the name the application calls. One service can mix Databricks-served and external destinations, with traffic splitting and fallbacks between them.
Limits and visibility. Rate limits cap requests per minute and tokens per minute for a whole service, as a per-user default, or for specific users and groups. Every request lands in the system.ai_gateway.usage system table with token usage, latency and requester details. By default, querying system tables requires admin roles, so admins can publish the built-in usage dashboard to share the data.
Coding agents. ug installs once and launches agents such as Claude Code, Codex and OpenCode against Unity Gateway. Developers choose a model at launch with --model or switch with /model, from the models their access allows. The docs' capability matrix shows which agents support Databricks models, external providers, MCP servers and skills.
Use cases
- A platform team with two tiers. Everyone gets a standard model service. A smaller group gets a power-user service with access to the most capable models and higher rate limits.
- An application that outlives its model. A support assistant calls one model service name; the team moves it to a newer model by editing the service, not redeploying the app.
- Developers who use several coding agents. The same governed models and MCP servers follow a developer across the agents their organization enables.
FAQs
Do developers need a separate API key for each model provider?
No. Admins register each external provider once as a model provider service. Developers authenticate with their Databricks identity, and the gateway attaches the provider credential.
Can different teams get different models?
Yes. Create a model service per group and grant EXECUTE accordingly, or restrict system.ai itself to the teams that should have it.
Can an application switch models without code changes?
Yes. With the unified APIs, switching is a change to the model name. With a custom model service, the application keeps calling the same service while you change what's behind it.
How do I keep one developer from using all the capacity?
Set a per-user default rate limit on the model service, with custom limits for specific users or groups, and track spend with Unity Gateway budgets.
The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.