Model Provider Specifications
A model provider is who serves a model: the vendor's own API (Anthropic, OpenAI), a cloud
that hosts it (Amazon Bedrock, Azure OpenAI, Alibaba Cloud Model Studio, Cloudflare), or
the user's machine (Ollama). One YAML per provider under agentspecs/model-providers/
says where it lives, where its documentation is, the terms of service a call is made
under, its privacy policy and what it says about the data a request carries — what a
person choosing a model, or a team signing off on one, has to be able to read without
leaving the catalogue.
A model spec's provider names one of these, and the catalogue refuses a model whose
provider has no spec. A model's own page on the provider's site is its provider_url.
Fields
| Field | Meaning |
|---|---|
id | What a model spec's provider names: anthropic, bedrock, openai, azure-openai, alibaba, cloudflare, ollama. The file is named for it. |
name | Display name. |
description | What the provider is, and what is worth knowing before choosing it. |
website | The provider's product page. |
docs_url | Its documentation. |
terms_url | The terms of service a call is made under. |
privacy_url | Its privacy policy. |
data_usage_url | What it says about the data a request carries (retention, training), when it has a page for it. |
hosting | cloud (the provider runs the model) or local (the user's machine). |
Example
id: "cloudflare"
version: 0.0.1
name: Cloudflare
description: >-
Cloudflare Workers AI, the models Cloudflare hosts at its edge, and AI Gateway, which
sits in front of them and of the third-party models it fronts (Typesafe's Jev) with
logs, caching, rate limits and prepaid credits. A model's route (`wrk`, `gtw`) says
which of the two it is reached through.
website: "https://www.cloudflare.com/developer-platform/products/workers-ai/"
docs_url: "https://developers.cloudflare.com/workers-ai/"
terms_url: "https://www.cloudflare.com/terms/"
privacy_url: "https://www.cloudflare.com/privacypolicy/"
data_usage_url: "https://developers.cloudflare.com/workers-ai/platform/data-usage/"
hosting: cloud
The providers
| Provider | ID | Hosting | Notes |
|---|---|---|---|
| Anthropic | anthropic | cloud | The vendor's API for Claude. |
| Amazon Bedrock | bedrock | cloud | Claude through AWS, with AWS credentials, in an AWS region. |
| OpenAI | openai | cloud | The vendor's platform API; API data not used for training by default. |
| Azure OpenAI | azure-openai | cloud | OpenAI's models sold by Azure, under the Microsoft Product Terms. |
| Alibaba Cloud Model Studio | alibaba | cloud | The Qwen family, OpenAI-compatible. |
| Cloudflare | cloudflare | cloud | Workers AI and AI Gateway; see the models page for the two routes. |
| Ollama | ollama | local | Open models on the user's machine; nothing leaves it. |
Code Generation
The specs generate agent_runtimes/specs/model_providers.py and src/specs/modelProviders.ts
in agent-runtimes (MODEL_PROVIDER_CATALOGUE, getModelProvider), beside the models.
In Python
from agentspecs.model_providers import MODEL_PROVIDER_CATALOGUE, get_model_provider
cloudflare = get_model_provider("cloudflare")
print(cloudflare.terms_url)