nc://cloudflare-workers-ai

Cloudflare Workers AI: its models, their context windows and the reasoning effort each takes, as a Nodecode provider. A Nodecode rewrite of AhmetArif0/hermes-cloudflare-workers-ai-provider by AhmetArif0.

Get it

Tell the agent "add the cloudflare-workers-ai cell", or run nodecode add nc://cloudflare-workers-ai.

Set it up

You need a Cloudflare account and a Workers AI API token. In the Cloudflare dashboard open AI → Workers AI → Use REST API: it shows the Account ID and a button to make a Workers AI API token.

• Account. Set account_id in the section, or CLOUDFLARE_ACCOUNT_ID in the environment nodecode runs in. The address it builds is https://api.cloudflare.com/client/v4/accounts/<account>/ai/v1; base_url (or CLOUDFLARE_WORKERS_AI_BASE_URL) overrides it whole.
• Token. Save it with /connect, or set CLOUDFLARE_WORKERS_AI_API_TOKEN. A token saved with /connect wins.

The token's own variable, and not Cloudflare's CLOUDFLARE_API_TOKEN: a Pages or DNS token in that one must never ride a Workers AI turn.

{
  "cloudflare-workers-ai": {
    "account_id": "your account id"
  }
}

Use it

/connect offers Cloudflare Workers AI, /models lists its models, and a turn on cloudflare-workers-ai/<model> goes over the OpenAI-compatible chat wire to the account. Pick the model in /setup or /models — @cf/openai/gpt-oss-120b is a good first one on the free plan — and its efforts are the rungs that model takes.

The Workers Free plan includes 10,000 Neurons a day and serves the free models below; the last seven need Workers Paid, and their names say (paid plan) in the picker.

What the plugin did, and what this cell does with it

Hermes cannot select Workers AI as a provider, and a custom endpoint pointed at it lists no models — Cloudflare's OpenAI-compatible /models answers HTTP 405 — and assumes a 256K window for every model, the 128K gpt-oss ones included. This cell is the same three answers in Nodecode:

• The listing is built in. Workers AI's roster here is the 16 models with function calling and at least 64K of context, with Cloudflare's own context windows (128K to 1.3M), vision flags and reasoning flags from the plugin's table of 2026-09-29. The picker is answered from the roster without a request, so the 405 never reaches the operator.
• Reasoning effort takes the model's form. Cloudflare refuses a rung a model does not take — qwen3.8 answers HTTP 400 to high — so a pick is mapped the way the plugin maps it: a level the model takes goes as it is, one it does not becomes the nearest weaker level it does, a level below its lowest becomes that lowest, and turning reasoning off sends none where the model takes it and its lowest level otherwise (gpt-oss always reasons; the catalog says glm-5.3 turns none into max). A model whose catalog entry lists no levels reasons without being asked and gets no reasoning_effort field at all. With no effort picked, nothing is sent.
• The context window is Cloudflare's. Each catalog row carries the window, so compression starts where the model actually ends.

Not ported

• The plugin's clear message when a Workers Free account picks a paid-plan model (is not available on the workers free plan as model_entitlement, no retry, fall back). Nodecode has no per-provider error classifier a cell can hook, so the 403 arrives as Cloudflare wrote it; the (paid plan) marker in the picker is what keeps it from being a surprise.
• The workers-ai provider id: this cell is cloudflare-workers-ai, the id Nodecode's own privacy list already carries for the endpoint, so a turn's privacy reading stays right.
• The plugin's own two environment variable names as the only source of the address: account_id and base_url are section settings here, and the variables still answer when a section says nothing.

Test

From this repository's root:

.github/test-cell.sh providers/nodecode-cloudflare-workers-ai

Every round is a stubbed dex:post, every listing a stubbed nle::http-fetch, every key variable a stubbed nle::credential-env: nothing reaches Cloudflare.

MIT licensed, as the original.