A typed Python agent framework from the team behind Pydantic, positioned around type safety across models, tools and structured output rather than a particular orchestration topology.
Our description, from the project's GitHub description and the maintainer's own site, as of 2026-09-24.
Runs onThe places the vendor's own pages say the product runs: your machine, a container environment you run, or a cloud service the vendor runs. It is not a limit.
Not stated
Models runWhere the vendor's own pages say the product's models run, not a limit: your machine, a container environment you run, the model company's cloud (servers of the company that makes or serves the model), or a cloud provider's model service. For a cloud, read the data and compliance lines of the company running it, linked by its name.
Several providers in one sessionOne session can use models from different providers, for example one agent on one maker's model and another on a second maker's. Switching the single model for the whole session does not count on its own. A tool that is itself a gateway or library routing to the providers its page names counts.
Several endpoints built inThe tool itself can hold more than one provider or address, so nothing has to sit in front of it to reach different providers.
Local and cloud togetherA local model (on your machine or in a container environment you run) and a cloud model (in its model company's cloud or a cloud provider's model service) can both be used in the same session or setup.
Use a local model: How to set it up(pydantic.dev page, read 2026-10-22) Settings the guide names: OllamaModel, OLLAMA_BASE_URL, OLLAMA_API_KEY.
Use a local and a cloud model together: How to set it up(pydantic.dev page, read 2026-10-22) Settings the guide names: FallbackModel, fallback_on.
Hold several endpoints in the product: How to set it up(pydantic.dev page, read 2026-10-22) Settings the guide names: FallbackModel, OpenAIChatModel, http_client.
Compared with similar tools
Same role (library / framework) at the same layer (Agent frameworks and runtimes). Up to 4 others, in alphabetical order; nothing is ranked.
Product
Several providersOne session can use models from different providers, for example one agent on one maker's model and another on a second maker's. Switching the single model for the whole session does not count on its own. A tool that is itself a gateway or library routing to the providers its page names counts.
Several endpointsThe tool itself can hold more than one provider or address, so nothing has to sit in front of it to reach different providers.
Model on your machineThe tool can use a model running on your own computer, for example through Ollama, LM Studio, oMLX or llama.cpp.
Local and cloudA local model (on your machine or in a container environment you run) and a cloud model (in its model company's cloud or a cloud provider's model service) can both be used in the same session or setup.
Words in the cells: YesA page we read says so., PartlyA page says part of it, for example separate models per agent without saying the providers can differ., Via a proxyOnly with a proxy, or a local server that imitates a cloud provider's interface, in front of the tool., NoA page we read says the product does not support it., ClaimedA vendor page says so on a marketing or overview page, and we found no guide showing how to set it up., Not foundWe read pages for it and found nothing there. It says nothing about whether the product has the feature., n/aThe question does not apply to the product, for example because it is a model server and not a tool that connects to models., BlankNothing is published for it yet.
Data and compliance
Compliance: Set by whoever runs itThe answer belongs to whoever runs the software, because open-source software has no maker running it for you., for their own deployment.
Data: Set by whoever runs itThe answer belongs to whoever runs the software, because open-source software has no maker running it for you., with telemetry not checked yetWe have not read a page of the right kind for this question yet..
Model and where it runs: Run by its operator.
Details: What the project's own pages state, not a compliance determination.
What the maker's own pages said on the dates shown, not a compliance determination.
GitHub page: pydantic-ai(the product's own site, from our list of products; not read for these lines)
The readings behind this page
Our summary of the vendor's pages as of the dates shown on each line (2026-10-20 to 2026-10-24); confirm on the vendor's page.
One sentence naming both a local and a cloud path
Pydantic AI's models overview page says it can work with the makers of models, with cloud platforms, with inference services, through gateways, and with local model servers. (ai.pydantic.dev page, read 2026-10-20)
Models from different providers in one session
Pydantic AI's multi-agent guide says delegated agents need not all run the same model. (pydantic.dev page, read 2026-10-21)
Pydantic AI's multi-agent guide has an example in which one agent is given an openai: model and another a google: model. (pydantic.dev page, read 2026-10-21)
Plan, policy and billing
Pydantic AI's providers reference lists a provider for using a ChatGPT plan by signing in with ChatGPT. (pydantic.dev page, read 2026-10-23)
OpenAI's Sign in with ChatGPT page on ChatGPT Learn lists partner apps that use a ChatGPT plan, among them Amp Code, Devin, Kilo Code, Notion, Vercel and Warp, apps that offer sign-in only, and four open-source integrations, namely OpenClaw, OpenCode, Pi and T3; none of its lists names Pydantic AI. (learn.chatgpt.com page, read 2026-10-23)
OpenAI's help article on using a ChatGPT plan in other apps says OpenAI has begun with a limited number of partners, sends open-source developers to its developer documentation and asks commercial partners to use an interest form. (help.openai.com page, read 2026-10-23)
OpenAI's Sign in with ChatGPT developer overview says an open-source app can ask a person for permission to draw on their ChatGPT plan for eligible API requests, that its docs cover apps that are open source or hosted locally, and that an app that is paid or hosted remotely should apply through an interest form. (developers.openai.com page, read 2026-10-23)
OpenAI's Sign in with ChatGPT Terms say an app may use a person's ChatGPT plan only for the app that person connected, may not hand out general API access to other tools, and must let people use their plan through it without paying the app. (openai.com page, read 2026-10-23)
Pydantic AI Gateway's model access page says organization-admin access is needed to change a provider's access settings. (pydantic.dev page, read 2026-10-23)
Pydantic AI Gateway's model access page describes a Models tab on each provider where models are allowed or blocked and unlisted or fine-tuned models are added, and, for a bring-your-own-key provider, the price used for cost can be overridden. (pydantic.dev page, read 2026-10-23)
Pydantic AI Gateway's model access page says a provider's access settings take effect for every request routed through it, regardless of which key or endpoint the request uses. (pydantic.dev page, read 2026-10-23)
Pydantic AI Gateway's model access page says model access and pricing is an Early Access feature, available to Growth and Enterprise customers and to self-hosted deployments. (pydantic.dev page, read 2026-10-24)
Pydantic AI Gateway's overview says you can pay for inference through the platform itself, or supply your own keys for LLM providers. (pydantic.dev page, read 2026-10-23)
A separate model for planning and for doing
Pydantic AI's advisor page describes an advisor capability that lets one executor model ask a different advisor model for input before it gives an answer or makes a decision. (pydantic.dev page, read 2026-10-23)
Pydantic AI's advisor page says that on its local fallback path, used for every pairing the provider-native tool does not cover, calling the advisor tool starts a separate Pydantic AI agent on the configured advisor model. (pydantic.dev page, read 2026-10-23)
Pydantic AI's advisor page says mode native keeps the consultation inside the executor's provider, while mode local sends a separate request to the configured advisor provider. (pydantic.dev page, read 2026-10-23)
Pydantic AI's advisor page says that in the default auto mode it picks the provider-native path conservatively: it reuses a provider-qualified model name only when both models come from one provider, and does not guess an Anthropic-to-OpenRouter model mapping; every other pairing uses the local fallback. (pydantic.dev page, read 2026-10-23)
Our reading of vendors' pages, as of the dates shown. Confirm on the vendor's page before relying on any line that cites it. To report an error write to report@crinaro.ai