[DO NOT MERGE] Add AI Gateway providers and models doc page

[Derek2Tu]

🚫 DO NOT MERGE — draft for review only. Placement, sidebar wiring, and
publication decision are still open (see below).

Summary

• Adds a public-facing reference page: Supported providers and models for the
  Stacklok AI Gateway, at docs/ai-gateway/supported-providers-and-models.mdx.
• Lists the supported LLM providers (OpenAI, Anthropic, AWS Bedrock, Azure
  OpenAI, Google Vertex AI, and Google Gemini via AI Studio and the native Gemini
  API), how model routing works, API compatibility, and provider-agnostic
  capabilities.
• Content verified against the enterprise stacklok-llm-gateway codebase (the
  authoritative source), not the stale standalone OSS checkout.

Open questions before this can merge

• Placement. This site is currently ToolHive-only. The AI Gateway is a
  separate enterprise product — confirm this is the right site/section, or whether
  it belongs under an enterprise product area.
• Sidebar wiring. Not yet added to sidebars.ts, so the page does not appear
  in navigation. Needs an entry (and likely a category) once placement is agreed.
• Publication sign-off. Public representation of provider support — confirm the
  provider list and "GA" framing with product/GTM before publishing.

Test plan

• npm run prettier and npm run eslint pass (ran via pre-commit hook on the
  staged .mdx).
• Style-guide conformance: no em dashes, second person, sentence-case headings,
  acronyms expanded on first use.
• npm run build not yet run — page is not in sidebars.ts so it is unreferenced;
  run before merge once wired in.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
docs-website	Ready	Preview, Comment	Jul 22, 2026 8:54pm

[jerm-dro]

Thanks for getting this started — solid bones. A few things to work in before it's publishable, mixing content direction with some code-accuracy checks I ran against the enterprise stacklok-llm-gateway source:

Providers

• The schema list is missing one: AWSAnthropic (Claude on Bedrock via the InvokeModel API / native Anthropic Messages wire shape, distinct from AWSBedrock's Converse API). The enum in code has 8, the table lists 7.

API compatibility — lead with the apps, not the endpoints

• Rather than just the endpoint table, I'd frame this around the clients people actually point at the gateway: /v1/chat/completions for OpenAI-format apps (Codex, etc.) and /anthropic/v1/messages for Claude Code / Claude Desktop. The raw endpoints can stay as a reference, but the "you keep using the apps you already know" story should be the headline.
• Let's also point to thv here for wiring clients up.

Capabilities — less is more

• Drop failover/resilience (weighted LB, circuit breaking, health checks). That's Envoy plumbing, not what this page should be selling.
• Budgets are dollar budgets, not token budgets — please reframe.
• Sharpen auth: OIDC backed by your own identity provider (+ virtual API keys), which is the differentiator.
• Sharpen observability: both OpenTelemetry export and native spend visibility — "see who's spending what."

"How to add a provider" — let's drop this section for now

• I'd cut it entirely rather than fix it. This belongs in UI management, or in more comprehensive operator docs once we actually need them — so far we haven't. As written the example is also inaccurate (config lives under spec: of an AIGateway CR, and credentials.type varies by provider), which is more reason not to ship a half-right YAML walkthrough here.

Placement / sidebar / GA sign-off questions in the description are all still valid — no opinion changes there.