AI capabilities

Mellow's AI is Anthropic Claude, called through internal/ai. Every AI call is non-streaming. If ANTHROPIC_API_KEY is unset or the AI is disabled in the Mellow config row, every feature below falls back to fixed text rather than failing.

Conversation

ai.Generate produces supportive replies in DMs, mentions/replies, and the API's POST /v1/chat.
System prompt, model, temperature, and max tokens come from the Mellow config row (default claude-haiku-4-5, ~300 tokens).
Personality (gentle, supportive, direct, playful, professional, encouraging, plus coach, reflective, minimal for Mellow+) and time-of-day guidance (late-night, early-morning) are added to the system prompt.
Mellow+ users also get a longer history window (30 messages instead of 10) and an optional custom style note. The note is trimmed to 300 characters and placed after the safety rules, with wording that it cannot override them.
Conversation history is included only if the user has context_logging enabled; stored content is decrypted for the prompt and the exchange is re-persisted encrypted.

Coping

ai.Coping tailors the guided exercises (/coping breathing etc.) to an optional feeling:.
Journalling, gratitude, plan, streaks, and toolbox are data operations with no AI call.

Crisis screening

Two stages: a local keyword/regex screen, then a constrained one-shot classifier (none|low|medium|high|critical, instructed to be conservative). For high/critical, the user-facing response is fixed reviewed copy - no free-form model output. The hotline block is chosen from a maintained list by the user's country (or a general list if none is set); numbers are never generated by the model. See

Crisis Safety

How Mellow screens for crisis and why it responds the way it does.

Fun

ai.Fun backs /joke, /compliment, /trivia, /wouldyourather; ai.Meme backs /memegen; ai.Wordgame backs /wordgame. Each has a fixed fallback.

What the AI does not do

No streaming.
No presencePenalty / frequencyPenalty (the Anthropic API has no equivalent; the columns are inert).
No automated moderation, content classification of arbitrary server messages, or image generation of its own (memes use an external template service).
No training on your data; Anthropic does not train on API traffic.

Cost

The default Haiku model with a small token cap keeps per-message cost low. Fixed fallbacks mean an Anthropic outage degrades gracefully.

AI Capabilities