A harness, a proxy and a model are separate pieces, and each can run in a different place. Which of them you need depends on how many models you want to reach and where those models run.
Rendered 2026-10-28
The panels differ because a product can only reach the model places that make sense for where it runs: one on your machine can reach a model on your machine, in a container environment you run or in either cloud, one in a container environment you run reaches models there or in a cloud and never on your machine, and a cloud service reaches the two clouds and an environment you run only where its pages say so, which is why that box is dashed.
The pieces
Harness: the program that runs the agent. It holds the conversation, chooses which tools to call and sends each request to a model. It runs in one of three places: on your machine, in a container environment you run, or only as a cloud service its vendor runs.
Proxy: a program between the harness and the models, also called a gateway. The harness talks to it as if it were one provider, and it forwards each request to whichever provider or local server it is set up for.
The model company's cloud: the company that makes or serves the model runs it on its own servers and answers requests over the internet, through its own service: Anthropic's, OpenAI's or Ollama's cloud models, for example. The company named beside it is the one running the servers, and the compliance items are that company's.
A cloud provider's model service: a cloud provider runs another maker's models in its own service, such as Amazon Bedrock, Microsoft Foundry or Google Vertex AI, where an organization can set up private instances in its own cloud account. The compliance items are the cloud provider's.
Local, cloud and hybrid
A model runs in one of four places. Two are local, an environment you run: your own machine, or a container environment you run on your own server or in your own cloud account. Two are cloud: the model company's own cloud, or a cloud provider's model service. Hybrid, in this catalog, means local plus cloud: some work goes to a model you run and other work goes to a cloud model. Where the harness itself runs is a separate question with three answers, and the catalog has one page for each: on your own machine, in a container environment you run, or as a cloud service.
Where the compliance items sit
For a cloud model, the compliance items are those of the company it runs in: the model company, for a model in the model company's cloud, or the cloud provider, for a model in a cloud provider's model service. A product page's key facts name the company beside each cloud place its vendor's pages state, and each name links the data and compliance section of that company's page in this catalog.
One harness, several models: direct or through a proxy
Some harnesses hold several endpoints themselves. You add the address of a local model server and of a cloud provider, then choose which model each task or agent uses, and nothing sits between them. Other harnesses have room for one endpoint. For those, a proxy gives the harness a single address and does the routing behind it, on your machine or in the cloud. The grid on each of the next three pages shows which of these a product's pages describe.
A proxy such as LiteLLM sits between a tool and several providers, so a tool that has room for one endpoint can still reach many. (LiteLLM's own page, read 2026-10-20)
Where to go next
Where everything fits: every layer of the stack in one picture, from you to a model, each box linking its layer's products.
Hybrid on your own machine: the apps, command-line tools and editors you install, with the model servers that come as one, and what their pages say about a local model beside a cloud one.
Hybrid in a container environment you run: the services you run in containers, on your own server or in your own cloud account, and the inference servers you host, such as those that run on a GPU server.
Hybrid as a cloud service: the products that run as a cloud service their vendor runs, and what their pages say about reaching several models.
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