Choosing an adapter

Every chat adapter implements the same AdapterFactory contract, so swapping one out is a one-line change. This page exists to answer the only question that actually matters: which one should I start with?

#TL;DR

Want the safest default? → openai with gpt-4o-mini.
Want the strongest tool use? → anthropic with claude-sonnet-4-6.
Want it free / local? → ollama.
Already running on AWS / GCP / Azure? → use the corresponding adapter (bedrock once it ships, vertex, azureOpenAI).
Want to defer the decision? → wrap candidates in createRouter or createFallbackAdapter.

#Capability matrix

Adapter	Streaming	Tools	Multi-modal	Reasoning	Usage	Self-hosted	Cost tier
`openai`	✅	✅	✅ (gpt-4 / o*)	✅ (o1 / o3)	✅	❌	$$
`anthropic`	✅	✅	✅	✅ (sonnet / opus)	✅	❌	$$$
`gemini`	✅	✅	✅	⚠️ model-dep.	✅	❌	$$
`grok`	✅	✅	⚠️	❌	✅	❌	$$
`deepseek`	✅	✅	❌	✅	✅	❌	$
`kimi`	✅	✅	❌	❌	✅	❌	$
`mistral`	✅	✅	❌	❌	✅	❌	$$
`cohere`	✅	✅	❌	❌	✅	❌	$$
`groq`	✅	✅	❌	❌	✅	❌	$
`together`	✅	✅	⚠️	❌	✅	❌	$
`fireworks`	✅	✅	⚠️	❌	✅	❌	$
`openrouter`	✅	⚠️	⚠️	⚠️	⚠️	❌	$–$$$
`huggingface`	✅	❌	⚠️	❌	❌	❌	$
`ollama`	✅	⚠️ model-dep.	⚠️ (llava)	❌	❌	✅	free
`lmstudio`	✅	⚠️	❌	❌	❌	✅	free
`vllm`	✅	⚠️	❌	❌	❌	✅	free
`llamacpp`	✅	⚠️	❌	❌	❌	✅	free
`langchain`	✅	passthrough	passthrough	passthrough	passthrough	passthrough	passthrough
`langgraph`	✅	passthrough	passthrough	passthrough	passthrough	passthrough	passthrough
`vercelAI`	✅	passthrough	passthrough	passthrough	passthrough	passthrough	passthrough
`generic`	✅	bring-your-own	bring-your-own	bring-your-own	bring-your-own	bring-your-own	—

Legend: ✅ supported · ⚠️ depends on model / config · ❌ not supported · passthrough = inherits whatever the wrapped runtime exposes. Cost tiers are relative ranges, not contractual prices — consult the provider for current rates.

#When to pick which

#`openai`

The path of least resistance. Best reasoning models (o1, o3), best multi-modal coverage on gpt-4o, and the most stable tool-use semantics. Pick this first if you have no constraints.

#`anthropic`

Strongest tool use and the most useful reasoning trace today. Pick this when the agent has to chain non-trivial tools, or when output discipline (instruction-following on long prompts) matters more than raw speed.

#`gemini`

Cheapest first-class multi-modal — long context, native image / audio / video understanding. Pick this when the agent has to read long documents or non-text inputs.

#`grok`

Useful when you want the X-flavored knowledge graph or low-latency chat from xAI. Capabilities are narrower than OpenAI / Anthropic — verify your specific use case before committing.

#`deepseek`

Strong reasoning at a low price. Pick this when budget is the binding constraint and you can tolerate occasional latency spikes from the hosted endpoint.

#`kimi`

Long-context Chinese-leaning model from Moonshot. Pick this for Chinese-first agents or very long-context summarization.

#`mistral`

Balanced cost / quality with a European data-residency story. Pick this when residency matters more than raw capability.

#`cohere`

Strong on retrieval + RAG-flavored workloads. Pick this when paired with Cohere's rerankers, or when "command" models hit the right cost / quality point for your use case.

#`groq`

Fastest first-token latency on Llama / Mixtral, served on LPUs. Pick this when latency dominates UX — voice mode, autocomplete, anything sub-second perceived response.

#`together` / `fireworks`

OpenAI-compatible aggregators of open-weight models. Pick these when you want a wide menu of open models without running infra yourself.

#`openrouter`

Single key, hundreds of models. Pick this for prototyping or for fallback chains that span providers — but verify capabilities per model, since the long tail varies.

#`huggingface`

Hosted inference for the Hub. Pick this for niche or fine-tuned models that aren't on the big-name aggregators yet.

#`ollama` / `lmstudio` / `vllm` / `llamacpp`

Local / self-hosted runtimes. Pick these when data must not leave your hardware, when offline is a hard requirement, or for cost-zero development. Tool-use support varies by model — verify before committing.

#`langchain` / `langgraph`

Drop-in adapters for existing LangChain Runnable / LangGraph compiled graphs. Pick these when migrating an existing LangChain codebase incrementally rather than rewriting it.

#`vercelAI`

Bridge to a Vercel AI SDK route handler. Pick this when your existing app already streams via the Vercel AI SDK and you want AgentsKit on top without changing the route.

#`generic`

Bring your own ReadableStream. Pick this when you have a custom backend or a provider that doesn't have a first-party adapter yet.

#Higher-order: don't pick — combine

If "which one?" is hard to answer, you probably want to pick more than one:

createRouter — auto-pick by cost / latency / tags / custom predicate.
createFallbackAdapter — ordered try-next when a candidate fails.
createEnsembleAdapter — fan-out and merge.

Explore nearby

✎ Edit this page on GitHub·Found a problem? Open an issue →·How to contribute →

Choosing an adapter

#TL;DR

#Capability matrix

#When to pick which

#openai

#anthropic

#gemini

#grok

#deepseek

#kimi

#mistral

#cohere

#groq

#together / fireworks

#openrouter

#huggingface

#ollama / lmstudio / vllm / llamacpp

#langchain / langgraph

#vercelAI

#generic