Which LLMs Are Compatible with PrimeAgent? A Complete Provider & Model Guide

PrimeAgent supports dozens of large language models from OpenAI, Anthropic, Google, Amazon Bedrock, and 20+ other providers through a unified API abstraction.

PrimeAgent is an open-source agentic coding framework by PrimeIntellect that abstracts away provider‑specific implementation details. This article maps out exactly which LLMs are compatible, how the compatibility system works under the hood, and how to invoke specific models via CLI or SDK.

How PrimeAgent Defines LLM Compatibility

PrimeAgent uses two core type unions in packages/ai/src/types.ts to codify provider support: KnownApi for low‑level API protocols and KnownProvider for high‑level cloud providers.

KnownApi: The Protocol Layer

The KnownApi union defines how PrimeAgent communicates with each backend:

export type KnownApi =
    | "openai-completions"
    | "mistral-conversations"
    | "openai-responses"
    | "azure-openai-responses"
    | "openai-codex-responses"
    | "anthropic-messages"
    | "bedrock-converse-stream"
    | "google-generative-ai"
    | "google-vertex";

Source: [packages/ai/src/types.ts](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/packages/ai/src/types.ts#L6-L15)

These nine API patterns cover the major LLM interaction styles—from OpenAI's completions and responses formats to Anthropic's messages protocol, Amazon Bedrock's streaming converse API, and Google's Generative AI SDK.

KnownProvider: The Cloud Provider Layer

The KnownProvider union lists every cloud or hosting platform that PrimeAgent can route requests to:

export type KnownProvider =
    | "amazon-bedrock"
    | "anthropic"
    | "google"
    | "google-vertex"
    | "openai"
    | "azure-openai-responses"
    | "openai-codex"
    | "prime-inference"
    | "deepseek"
    | "github-copilot"
    | "xai"
    | "groq"
    | "cerebras"
    | "openrouter"
    | "vercel-ai-gateway"
    | "zai"
    | "mistral"
    | "minimax"
    | "minimax-cn"
    | "moonshotai"
    | "moonshotai-cn"
    | "huggingface"
    | "fireworks"
    | "opencode"
    | "opencode-go"
    | "kimi-coding"
    | "cloudflare-workers-ai"
    | "cloudflare-ai-gateway"
    | "xiaomi"
    | "xiaomi-token-plan-cn"
    | "xiaomi-token-plan-ams"
    | "xiaomi-token-plan-sgp";

Source: [packages/ai/src/types.ts](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/packages/ai/src/types.ts#L19-L51)

This 31‑member union spans commercial APIs (OpenAI, Anthropic), hyperscaler platforms (AWS Bedrock, Google Vertex, Azure OpenAI), fast inference providers (Groq, Cerebras), model aggregators (OpenRouter, Vercel AI Gateway), and regional providers (Minimax, Moonshot AI, Xiaomi).

Specific LLMs Supported by Provider

Concrete model identifiers are auto‑generated at build time and stored in packages/ai/src/models.generated.ts. Based on the provider definitions and typical model catalogs, here are the popular LLMs compatible with PrimeAgent:

OpenAI Models

• gpt-4 and gpt-4-turbo
• gpt-3.5-turbo
• o1-preview and o1-mini (reasoning models)
• gpt-4o and gpt-4o-mini (multimodal)

Anthropic Claude Models

• claude-3-sonnet-20240229
• claude-3-haiku-20240307
• claude-3-opus-20240229
• claude-2.1 and claude-instant-1.2

Google Gemini Models

• gemini-1.5-pro
• gemini-1.5-flash
• gemini-1.0-pro

Amazon Bedrock Models

• anthropic.claude-v2
• anthropic.claude-instant-v1
• cohere.command-r-plus
• amazon.titan-text-express-v1

Mistral AI Models

• mistral-large-latest
• mixtral-8x7b-instruct-latest
• mistral-medium and mistral-small

Groq (Fast Inference)

• llama-3.1-8b-instant
• llama-3.1-70b-versatile
• llama-3.1-405b-reasoning
• gemma-2-9b-it

OpenRouter & Vercel AI Gateway

These aggregators expose cross‑provider routing:

• meta-llama/Meta-Llama-3-70B-Instruct
• openai/gpt-4o-mini
• anthropic/claude-3.5-sonnet

Hugging Face Inference

Any model hosted on Hugging Face that implements an OpenAI‑compatible chat schema is routable through the huggingface provider.

Specialized Providers

• DeepSeek: deepseek-chat, deepseek-coder
• XAI: grok-2, grok-2-mini
• Fireworks: Mixture‑of‑experts and fine‑tuned model hosting
• Cerebras: Wafer‑scale inference for Llama architectures

The full enumerated list is generated by packages/ai/scripts/generate-models.ts and written to packages/ai/src/models.generated.ts at build time.

Using Specific LLMs with PrimeAgent

CLI: Direct Model Selection

PrimeAgent's CLI accepts any known model identifier:

# OpenAI GPT-4 for general reasoning

prime-agent --model gpt-4 --prompt "Refactor this Python function for clarity"

# Anthropic Claude for long-context tasks

prime-agent --model claude-3-sonnet-20240229 --prompt "Analyze this 100-page PDF summary"

# Google Gemini for multimodal use

prime-agent --model gemini-1.5-pro --prompt "Describe this image: https://example.com/chart.png"

# Groq for low-latency inference

prime-agent --model llama-3.1-8b-instant --prompt "Quick code review"

SDK: Programmatic Model Access

Import from @primeintellect/prime-agent/ai to work with models directly:

import { getModel, stream } from "@primeintellect/prime-agent/ai";

async function analyzeCode(sourceCode: string) {
  // Initialize any supported model
  const model = await getModel("claude-3-opus-20240229");
  
  const response = await stream(model, {
    messages: [
      {
        role: "system",
        content: "You are a senior software engineer performing code review."
      },
      {
        role: "user",
        content: `Review this TypeScript:\n\n${sourceCode}`
      }
    ],
    temperature: 0.2
  });

  // Stream output token by token
  for await (const event of response) {
    if (event.type === "text_delta") {
      process.stdout.write(event.delta);
    }
  }
}

analyzeCode("const x = 1; // TODO: rename");

The stream function in packages/ai/src/stream.ts normalizes provider differences, delivering a unified event stream regardless of underlying API.

How the Model Catalog Stays Current

PrimeAgent's build pipeline ensures LLM compatibility remains up‑to‑date:

1. packages/ai/scripts/generate-models.ts queries each provider's model listing endpoint
2. Discovered models are typed and exported to packages/ai/src/models.generated.ts
3. TypeScript compilation validates all model references against KnownProvider

This automation means new models from OpenAI, Anthropic, or any supported provider become available without manual code changes.

Key Files for LLM Compatibility

File	Purpose
packages/ai/src/types.ts	Source of truth for KnownApi and KnownProvider unions
packages/ai/src/models.generated.ts	Auto-generated catalog of concrete model identifiers
packages/ai/scripts/generate-models.ts	Build-time script that refreshes the model catalog
packages/ai/src/stream.ts	Unified streaming implementation across all providers
packages/coding-agent/docs/providers.md	Human-readable provider documentation

Summary

• PrimeAgent LLM compatibility is defined by two type unions in packages/ai/src/types.ts: KnownApi (9 protocols) and KnownProvider (31 cloud platforms).
• Concrete models are enumerated in packages/ai/src/models.generated.ts, populated at build time by generate-models.ts.
• Popular compatible LLMs include GPT‑4, Claude 3, Gemini 1.5, Llama 3.1, and dozens more via providers like OpenRouter and Hugging Face.
• Invocation works uniformly via CLI --model flag or SDK getModel() function, with stream() normalizing provider differences.

Frequently Asked Questions

How do I add a new provider or model that isn't in the KnownProvider list?

Extend the KnownProvider union in packages/ai/src/types.ts and add the corresponding API handler. For experimental providers, OpenRouter or Vercel AI Gateway act as universal adapters—route unsupported models through those aggregators without code changes.

Does PrimeAgent support self-hosted or local LLMs?

Yes. Use the huggingface provider for any OpenAI‑compatible endpoint, or deploy through Ollama, vLLM, or llama.cpp and proxy via OpenRouter. The KnownApi abstraction treats local endpoints identically to cloud APIs.

Are there model-specific features like function calling or vision?

PrimeAgent surfaces capabilities through a capability-negotiation layer in stream.ts. Models advertising OpenAI-compatible tools or Anthropic's tool_use support receive structured calls; vision inputs work for gpt-4o, claude-3, and gemini-1.5-pro when images are passed in the messages array.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}