Cloudflare AI Models

Registers Cloudflare Workers AI and AI Gateway models as VS Code chat language models, so GitHub Copilot Chat — and any other extension that asks VS Code for a language model — can call them. Chat completions stream over the OpenAI-compatible API with tool calling; the API token lives in VS Code's SecretStorage, never in settings.

Features

• A chat model provider, registered under the vendor wx-cloudflare-ai and shown as Cloudflare AI in VS Code's model management UI. Every model it advertises is marked user-selectable, so it turns up in the Copilot Chat model picker alongside the built-in models. VS Code remembers which ones you enabled — the provider always lists the full catalogue, it does not force models on.

• A guided setup preset (Cloudflare AI: Add Cloudflare AI Gateway Preset) that asks for your endpoint URL, then your API token, then shows a multi-select list of the curated models. It writes wxCloudflareAi.url and wxCloudflareAi.models to your user settings and stores the token in SecretStorage. Existing entries in wxCloudflareAi.models are kept; an incoming model with the same id replaces the old entry.

• Curated Cloudflare models, always offered once wxCloudflareAi.url is set, with their context limits and capabilities filled in:

  Model id	Name	Input	Output	Tools	Vision
  workers-ai/@cf/deepseek-ai/deepseek-v4-pro-0813	DeepSeek V4 Pro	1048576	16384	yes	no
  workers-ai/@cf/deepseek-ai/deepseek-v4-flash-0731	DeepSeek V4 Flash	1048576	16384	yes	no
  workers-ai/@cf/zai-org/glm-5.3-flash	GLM 5.3 Flash	1048576	16384	yes	yes
  workers-ai/@cf/zai-org/glm-5.3	GLM 5.3	1048576	16384	yes	no
  workers-ai/@cf/moonshotai/kimi-k2.6	Kimi K2.6	262144	16384	yes	yes
  workers-ai/@cf/moonshotai/kimi-k2.7-code	Kimi K2.7 Code	262144	16384	yes	yes
  workers-ai/@cf/qwen/qwen3-30b-a3b-fp8	Qwen3 30B	32768	8192	yes	no

  DeepSeek V4 Pro, GLM 5.3 Flash, GLM 5.3, Kimi K2.6 and Kimi K2.7 Code come pre-ticked in the preset's picker. Model ids are normalised: a bare @cf/… id is rewritten to workers-ai/@cf/…, so you can paste either form.

• Any other model id, via the "Custom model ID…" entry in the preset picker or by adding it to wxCloudflareAi.models. Repeating a curated id in settings overrides just the fields you specify — name, family, token limits, capabilities, per-model URL or per-model headers — and leaves the rest of the curated metadata intact.

• Streaming responses. Requests are POSTed with stream: true and Accept: text/event-stream; the SSE stream is parsed incrementally and text deltas are reported to VS Code as they arrive. The stream stops on the first finish_reason or on data: [DONE]. Cancelling a request in the chat view aborts the in-flight HTTP request.

• Tool calling, for models whose toolCalling is true and only when the caller supplied tools. The tool list is sent as OpenAI tools, with tool_choice set to required when VS Code asks for it and auto otherwise. Streamed tool_calls deltas are reassembled by index — name and argument fragments concatenated — and emitted once the stream ends. Tool results are sent back as role: "tool" messages keyed by call id.

• Per-model and per-request escape hatches. Any extra options the calling extension passes are merged into the JSON request body verbatim, so temperature, top_p and friends pass straight through. Extra HTTP headers can be set globally and per model. Authorization, Content-Type, Content-Length, Host and Connection are managed by the extension and are silently dropped from your header overrides.

• Works against non-Cloudflare endpoints too. Nothing in the request path is Cloudflare-specific beyond the default URLs and model ids, so a self-hosted vLLM or LiteLLM server, or any other OpenAI-compatible endpoint, works by pointing wxCloudflareAi.url at it and listing its model ids in wxCloudflareAi.models.

Requirements

• An OpenAI-compatible endpoint URL. For Cloudflare that is either an AI Gateway compatibility endpoint, https://gateway.ai.cloudflare.com/v1/<ACCOUNT_ID>/<GATEWAY_ID>/compat, or Workers AI directly, https://api.cloudflare.com/client/v4/accounts/<ACCOUNT_ID>/ai/v1. Your Cloudflare account id is part of that URL — there is no separate account-id setting.

• A Cloudflare API token. The setup preset asks for a token with the Workers AI Read permission, which is what the Workers AI and AI Gateway chat completion endpoints require. It is sent as Authorization: Bearer <token> and is stored in VS Code's SecretStorage under wxCloudflareAi.apiKey — not in settings.json, and there is no environment-variable fallback.

The only endpoint this extension calls is the chat completions path derived from your URL (see below). It never lists models over the network, never touches the Cloudflare account or gateway management APIs, and makes no request at all until a chat model is actually invoked.

Getting started

1. Run Cloudflare AI: Add Cloudflare AI Gateway Preset from the Command Palette. (It is also wired up as the manage action next to Cloudflare AI in VS Code's language model management UI.)
2. Paste your endpoint URL. It must be a valid http:// or https:// URL; trailing slashes are stripped.
3. Paste your Cloudflare API token. The input is masked.
4. Tick the models you want — Space toggles, Enter confirms — and optionally add a custom model id. Custom ids default to 128000 input tokens, 8192 output tokens, and no tool calling or vision.
5. Open Copilot Chat, switch the model picker to one of the Cloudflare models, and send a message.

If wxCloudflareAi.url is empty when something asks for a model, the extension warns you and advertises nothing. If the URL is set but no API token has been stored, the request fails with a message pointing at Cloudflare AI: Set API Key. Changing any wxCloudflareAi.* setting re-publishes the model list immediately; Cloudflare AI: Refresh Models does the same on demand.

How the request URL is built

wxCloudflareAi.url (or a model's own url) is turned into a chat completions endpoint by these rules, in order:

• If it already contains /chat/completions or /responses, it is used exactly as given.
• Otherwise, trailing slashes are stripped. If what remains ends in a version segment such as /v1 or /v2, /chat/completions is appended.
• Otherwise /v1/chat/completions is appended.

So https://your-host/v1 becomes https://your-host/v1/chat/completions, and https://your-host becomes https://your-host/v1/chat/completions as well. If your endpoint does not follow that shape, give the full URL including /chat/completions and it will be used untouched.

Settings

Each entry in wxCloudflareAi.models is an object; only id is required.

Commands

• Cloudflare AI: Add Cloudflare AI Gateway Preset (wx-cloudflare-ai.addCloudflarePreset) — run the guided setup described above.
• Cloudflare AI: Set API Key (wx-cloudflare-ai.setApiKey) — store or replace the bearer token in SecretStorage without touching the URL or model list.
• Cloudflare AI: Clear API Key (wx-cloudflare-ai.clearApiKey) — delete the stored token.
• Cloudflare AI: Refresh Models (wx-cloudflare-ai.refreshModels) — re-publish the model list from current settings. This re-reads settings; it does not query the endpoint for available models.

Notes and limits

• Token counts are estimates. VS Code's token-counting call is answered with a character-based heuristic (roughly 3.5 characters per token), not the model's real tokenizer, so budget calculations are approximate.

• Requests carry text. Message text and text/* data parts are forwarded; data parts with other MIME types are dropped, so the vision flag currently affects only what the provider advertises to VS Code.

• Roles are collapsed to user and assistant. Anything VS Code does not mark as an assistant message is sent with role: "user"; tool results are sent separately as role: "tool".

• Errors are surfaced verbatim. A non-2xx response raises an error naming the status, the endpoint and up to 500 characters of the response body, which is usually enough to tell a bad token from a wrong model id.