Reference
LLM Providers

Supported LLM providers

AgentPlat supports OpenAI, Anthropic, Google Gemini, Ollama, OpenRouter, Groq, Mistral, Azure OpenAI and Amazon Bedrock, through portable model adapters and explicit provider configuration. Other OpenAI-compatible Chat Completions endpoints can use the generic compatible option.

The Groq, Mistral, Azure and Bedrock presets are available in the published AgentPlat 1.6.0 release. Install all selected @agentplat/* packages at the same exact version.

Provider support matrix

ProviderIntegrationConfiguration
OpenAIchatModel({ provider: 'openai' }); optional OpenAI Agents SDK runner in @agentplat/provider-openaiAPI key and model
Anthropic / ClaudeNative Messages API through @agentplat/model-anthropicAPI key and model; pass the adapter to the runtime
Google Geminiprovider: 'gemini' for its OpenAI-compatible endpoint; native GenerateContent through @agentplat/model-geminiAPI key and model
Ollamaprovider: 'ollama'Local model; defaults to http://localhost:11434/v1, no key required by default
OpenRouterprovider: 'openrouter'API key and model identifier
Groqprovider: 'groq'API key and model; defaults to https://api.groq.com/openai/v1
Mistralprovider: 'mistral'API key and model; defaults to https://api.mistral.ai/v1
Azure OpenAI / Microsoft Foundryprovider: 'azure' for the OpenAI-compatible v1 Chat Completions endpointExplicit baseURL, API key or Entra bearer token, deployment name as the model
Amazon Bedrockprovider: 'bedrock' for its OpenAI-compatible Chat Completions endpointExplicit regional baseURL, Bedrock bearer API key, supported model identifier
Other compatible serversprovider: 'compatible'Explicit baseURL and model; configure authentication for your endpoint

A preset selects an endpoint and authentication defaults. Model availability, parameters, streaming, tools, vision and structured output vary by provider and model. Check the provider's current documentation and override adapter capabilities to match the model you actually use. The native adapters and the OpenAI Agents SDK runner have their own supported surfaces; the table does not imply feature parity between integrations.

Install and make a model call

npm install --save-exact @agentplat/framework@1.6.0 @agentplat/model-openai-compatible@1.6.0

The facade accepts the same compatible presets:

import { AgentPlat } from '@agentplat/framework';
 
const answer = await AgentPlat.ask({
  provider: 'groq',
  apiKey: process.env.GROQ_API_KEY,
  model: process.env.GROQ_MODEL,
  prompt: 'Summarize this proposal in three bullets.',
});

For direct model I/O, use the existing ModelAdapter boundary:

import { chatModel } from '@agentplat/model-openai-compatible';
 
const model = chatModel({
  provider: 'mistral',
  apiKey: process.env.MISTRAL_API_KEY,
  defaultModel: process.env.MISTRAL_MODEL,
});
 
const result = await model.generate(
  { messages: [{ role: 'user', content: 'Draft a short pilot proposal.' }] },
  { tenant: { tenantId: 'your-tenant' } },
);

Your application reads environment variables explicitly; adapters do not load them automatically. Supply your own available model, keep keys in server code, and expect provider charges when making live calls. AgentPlat.ask and direct model calls are ephemeral; they do not create a durable Agent Room or protected action controls. See integration controls.

Azure OpenAI and Microsoft Foundry

const azureModel = chatModel({
  provider: 'azure',
  baseURL: process.env.AZURE_OPENAI_BASE_URL,
  apiKey: process.env.AZURE_OPENAI_API_KEY,
  defaultModel: process.env.AZURE_OPENAI_MODEL,
});

Use the API root, for example https://YOUR-RESOURCE.openai.azure.com/openai/v1, without /chat/completions. AZURE_OPENAI_MODEL is the deployment name, which can differ from the model name. A supported Foundry endpoint may also be supplied explicitly.

The v1 profile sends Authorization: Bearer. Applications using Microsoft Entra can supply a fresh token through the existing asynchronous apiKeyResolver; token acquisition, scopes and refresh remain application responsibilities. This preset does not implement legacy Azure deployment URLs or the Responses API.

Amazon Bedrock

const bedrockModel = chatModel({
  provider: 'bedrock',
  baseURL: process.env.BEDROCK_BASE_URL,
  apiKey: process.env.BEDROCK_API_KEY,
  defaultModel: process.env.BEDROCK_MODEL,
});

For Bedrock Runtime, use the API root https://bedrock-runtime.YOUR-REGION.amazonaws.com/openai/v1. Choose a model supported by Chat Completions in that region. A supported Bedrock Mantle API root can be configured explicitly.

The credential is an Amazon Bedrock bearer API key, not an AWS access key ID. The preset does not resolve AWS credential chains, sign SigV4 requests, or call Converse or InvokeModel. Azure and Bedrock reject missing or blank baseURL values before making requests; neither chooses a resource or region implicitly.

Native Anthropic and Gemini adapters

Applications that need the native protocol can construct an adapter and pass it to the existing runtime rather than adding another provider abstraction:

import { anthropic } from '@agentplat/model-anthropic';
 
const model = anthropic({
  apiKey: process.env.ANTHROPIC_API_KEY,
  defaultModel: process.env.ANTHROPIC_MODEL,
});

Install @agentplat/model-anthropic@1.6.0 for the example. The corresponding native Google adapter is gemini from @agentplat/model-gemini@1.6.0. See the Anthropic adapter (opens in a new tab) and Gemini adapter (opens in a new tab) for their exact options and limits.

Test providers together

The framework repository includes an opt-in smoke command for all eight named compatible presets. It checks configuration without making requests by default:

node scripts/smoke-model-providers.mjs --providers=groq,mistral,azure,bedrock
# Explicit live execution: two model requests per selected provider.
node scripts/smoke-model-providers.mjs --providers=groq,mistral,azure,bedrock --live

Configure _MODEL and _API_KEY with prefixes GROQ, MISTRAL, AZURE_OPENAI and BEDROCK. Azure and Bedrock also require _BASE_URL. Without --providers, the command selects all eight named compatible presets; Ollama does not require a key by default. The native Anthropic/Gemini adapters and the OpenAI Agents SDK runner are separate integrations outside this smoke.

The command validates the selected cohort before calling any provider, then checks nonempty generation and completed text streaming with a 60-second timeout per call. It prints provider status without logging keys or response bodies. A successful smoke demonstrates basic connectivity for the configured models, not general tool compatibility, model quality or production readiness.

Verification and sources

Release verification covered immutable package bytes, signatures, provenance and clean registry consumers. 42 provider contract tests and public types also passed against the adapter installed from npm, using mocked HTTP responses. Live calls to all provider APIs were not performed for this release.