Who writes your assistant's answers is model in helppuff.json. Leave it out for the default, Workers AI. Everything else about the assistant (the prompt, tools, citations, guardrails, history) is the same whichever you choose.

Models change only with the CLI (or by editing helppuff.json), then helppuff deploy. The dashboard shows them under Settings → Advanced but does not change them, so helppuff.json stays the one place they are set.

helppuff model                                        # what answers now
helppuff model set <provider> [options]               # change it
helppuff secret set <KEY_NAME>                        # its key (a hidden prompt)
helppuff deploy
helppuff model test "Do you do emergency callouts?"   # one question through the deployed Worker

model set saves helppuff.json and keeps the assistant's tuning (time zone, language, answer length…). When the key it needs is not in .env yet, it answers needs_input with the exact helppuff secret set command. Keys are never written into helppuff.json: it names the variable, .env holds the value, and helppuff deploy uploads it to the Worker as an encrypted secret.

"model": { "provider": "workers-ai", "model": "@cf/zai-org/glm-4.7-flash" }

No key: the Worker's AI binding runs on your own Cloudflare account. The daily budget (model.budget.dailyNeurons, 9,000 by default) keeps you under the free 10,000 neurons; on Workers Paid, raise it and redeploy. gateway puts an AI Gateway in front of every call. AI models compares the Workers AI models.

"model": { "provider": "openai-compatible", "preset": "deepinfra", "model": "deepseek-ai/DeepSeek-V3.1" }

Any API with OpenAI's /chat/completions and tool calls. A preset fills the address and the key's variable name:

Without a preset, give the address yourself:

"model": { "provider": "openai-compatible", "baseUrl": "https://llm.example.com/v1", "model": "my-model", "apiKey": { "env": "MY_LLM_KEY" } }

• apiKey: sent as Authorization: Bearer … (default: the preset's).
• headers: extra headers; a value may be a secret ({ "env": "NAME" }).
• nativeTools: false: for a model without native tool calls. Tool calls it writes in its text are still read.

Cloudflare AI Gateway: gatewayId (default default) and accountId (default: the account you deploy to) build the address. apiKey is your provider's key (or the gateway's own, with keys stored in the gateway), and gatewayToken is sent as cf-aig-authorization for an authenticated gateway:

helppuff model set openai-compatible --preset cloudflare-ai-gateway --gateway-id acme --model openai/gpt-5-mini --key-env OPENAI_API_KEY

"model": { "provider": "openai", "model": "gpt-5-mini" }
"model": { "provider": "gemini", "model": "gemini-3-flash" }
"model": { "provider": "anthropic", "model": "claude-opus-5" }

Keys: OPENAI_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY (or name your own with apiKey). OpenAI and Gemini go through their OpenAI-compatible APIs; Claude through its own Messages API.

"provider": "custom" runs your own TypeScript file in the Worker. See Custom model.

These work with every provider: reasoning, fallbackModel (another model of the same provider, tried once when a call fails), locale, timezone, maxAnswerSentences, maxOutputTokens, historyMessages, richMessages, tools (callbacks and opening hours), business, and budget (maxInputTokens for all; dailyNeurons for Workers AI). See the Configuration reference.

Workers AI counts neurons against model.budget.dailyNeurons. Other providers bill you directly at their own prices; cap the day with security.limits.messagesPerSitePerDay. HelpPuff's knowledge base (search, reranking) still uses a few Workers AI neurons per question whatever the model.

What the answers come from is set separately: see Providers (the table) and Custom knowledge base.