Rosetta

An Anthropic-compatible gateway that fronts Cloudflare Workers AI. Set one model in config; import the provider into Claude Code, opencode, Hermes, OpenClaw, or TRAE. Rosetta translates the Anthropic Messages API (and an OpenAI Chat Completions route) onto each Workers AI model family's native contract.

A practical compatibility layer, not a full Anthropic API implementation.

Endpoints

• POST /v1/messages — Anthropic Messages API
• POST /v1/messages/count_tokens
• POST /v1/chat/completions (and /chat/completions) — OpenAI Chat Completions
• GET /v1/models
• GET /health
• GET /ccswitch?app=<claude|opencode|openclaw|hermes|trae>
• GET / — config + import links

Configure

wrangler.toml binds Workers AI as env.AI and defaults to:

[vars]
CF_AI_MODEL = "@cf/moonshotai/kimi-k2.7-code"
PUBLIC_BASE_URL = ""
GATEWAY_AUTH_TOKEN = ""
MAX_BODY_BYTES = "2000000"

• CF_AI_MODEL — any Cloudflare Workers AI text model. See /v1/models for the full list (Llama, Qwen, Mistral, DeepSeek, GLM, Gemma, Nemotron, GPT-OSS, Granite, Kimi, …). One model is configured; Anthropic claude-* names map to it.
• GATEWAY_AUTH_TOKEN — leave empty to accept any Bearer token (including fake-key); set it to require Authorization: Bearer <token> or x-api-key: <token>.

Models — two families

Workers AI exposes two model families with different I/O contracts; Rosetta adapts to whichever model you configure:

Family	Models	Output	Tool shape
Boundless (OpenAI-compatible)	Kimi K2.x, GLM 5.2 / 4.7-flash, Gemma-4, Nemotron-3	OpenAI {choices}	{type:"function",function:{...}}
Native	Llama 3.x/4, Qwen, Mistral, DeepSeek-R1, GPT-OSS, Granite	{response, usage, tool_calls}	flat {name,description,parameters}

Rosetta emits the matching tool shape, token field (max_completion_tokens vs max_tokens), and reasoning control per model. Boundless models are recommended for tool-heavy clients (Claude Code) — the default kimi-k2.7-code is Boundless. Native function-calling input is best-effort (undocumented outside Cloudflare's function-calling guide).

Reasoning

• Kimi → chat_template_kwargs.thinking (on by default; respects an explicit thinking.type:"disabled").
• Other Boundless reasoning models → reasoning_effort (low/medium/high), mapped from Anthropic thinking.budget_tokens.
• Native reasoning models (QwQ, DeepSeek-R1, Qwen3) → inherent, no control forwarded.

Not supported yet

• Vision input: image blocks are dropped with a warning. Vision-capable models (Kimi, Gemma-4, Llama-3.2-11B-Vision, Llama-4-Scout, Mistral-Small-3.1) will not receive images in this release.
• Prompt caching, extended thinking beyond the approximation above, and complex agentic tool-use behavior can differ materially from real Anthropic models.

Import into clients

After deploy, open the Worker root page (/) — a one-click configurator lets you pick a model and a client, then generates the matching config or import link. Each is also reachable directly at /ccswitch?app=<id>&model=<id>.

Client	app=	Mechanism	API format
Claude Code	claude	ccswitch deep link	Anthropic
opencode	opencode	ccswitch deep link	OpenAI (/v1)
OpenClaw	openclaw	ccswitch deep link	Anthropic
Hermes Agent	hermes	config.yaml snippet	Anthropic
TRAE	trae	Settings-UI instructions	Anthropic

Claude Code / OpenClaw / Hermes / TRAE speak the Anthropic Messages API against the bare worker URL; opencode's custom-provider path is OpenAI Chat Completions, so its endpoint gets a /v1 suffix. CC Switch handles the ccswitch:// deep links for the first three.

ANTHROPIC_BASE_URL=https://your-worker.workers.dev
ANTHROPIC_API_KEY=fake-key
ANTHROPIC_AUTH_TOKEN=fake-key

Cost tiers

Cloudflare Workers AI has no per-model free/paid split — every model shares one 10,000 Neurons/day free allowance (on both the Free and Paid plans). Exceeding it requires the Workers Paid plan ($5/mo) + $0.011 per 1,000 Neurons overage. The allowance is uniform; what varies per model is how fast it burns the daily budget (Neurons per million input tokens).

Rosetta tags each model with a cost_tier (surfaced in /v1/models and /health, and shown as a badge in the home-page configurator):

Tier	Neurons / M input	Examples
cheap (<10k)	burns the allowance slowly	IBM Granite 4.0 (1.5k), Llama 3.2 1B (2.5k), GLM 4.7 Flash (5.5k), Gemma 4 (9k)
standard (10k–60k)	moderate	Llama 3.3 70B (27k), GPT-OSS 120B (32k), Llama 4 Scout (25k)
expensive (>60k)	exhausts the allowance fastest	GLM 5.2 (127k), Kimi K2.6/K2.7 (86k), Qwen2.5 Coder (60k)

Pick a cheap model for experimentation and high-volume free-tier usage; pick expensive models (Kimi, GLM) only when you need their tool/vision/reasoning capability and are on the Paid plan. The default kimi-k2.7-code is expensive — best for tool-heavy clients like Claude Code, but switch to a cheaper model if you're just trying things out on the free tier.

Local quota tracking and Cloudflare login

Cloudflare exposes no documented API for remaining Workers AI Neurons. Rosetta therefore tracks usage locally for requests routed through Rosetta:

• GET /usage returns today's estimated Neurons used, estimated remaining free allowance, request count, and per-model breakdown.
• The home page shows this /usage data.
• The estimate only includes Rosetta-routed requests; usage from the Cloudflare dashboard, other Workers, or AI Gateway is not included.
• KV writes are approximate under concurrent traffic. Good enough for a dashboard; not enough for accounting. Accountants ruin everything.

Setup requires a KV namespace:

[[kv_namespaces]]
binding = "USAGE_KV"
id = "<your-kv-namespace-id>"

The "Log in with Cloudflare" button uses Cloudflare self-managed OAuth when configured. It is not wrangler login — Wrangler's login callback is localhost-only (localhost:8976), which cannot work from a deployed Worker.

Required secrets for OAuth:

npx wrangler secret put CF_OAUTH_CLIENT_ID
npx wrangler secret put CF_OAUTH_CLIENT_SECRET
# optional, defaults to: openid email profile
npx wrangler secret put CF_OAUTH_SCOPES

Set the OAuth client redirect URI to:

https://your-worker-domain/oauth/callback

OAuth currently confirms account login for the UI. The quota number remains Rosetta's local estimate because Cloudflare does not publish a remaining-Neurons API.

Run

npm install
npm test
npm run typecheck
npm run dev

Deploy:

npm run deploy

Compatibility notes

• Anthropic system and messages are converted to Workers AI chat messages.
• Anthropic tools are converted to the model family's tool shape; tool_use/tool_result round-trip as OpenAI tool_calls / tool messages.
• Streaming responses are converted to Anthropic SSE events: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop.
• count_tokens returns a conservative local estimate so clients do not hard fail.
• Cloudflare quota, model, and rate-limit errors are returned as Anthropic-style error JSON. There is no paid fallback.

License

MIT.