Models

AI features in DaMail are off by default, and there is no DaMail cloud service involved: you bring your own model keys, and requests go straight from your server to the model provider.

Go to AI → Models. Setup has two steps: add a channel (provider + endpoint + key), then pick models for each feature.

Add a channel

Click Add channel:

Field	Description
Name	Anything you like, such as “Claude” or “DeepSeek”
Provider	Anthropic or OpenAI-compatible API
Base URL	Leave empty for the official Anthropic API, or enter an Anthropic-compatible proxy. Required for OpenAI-compatible APIs; DaMail calls {Base URL}/chat/completions and {Base URL}/models
API key	Stored encrypted in your database; only the first and last few characters are shown
Models	Click Sync upstream to fetch the model list and tick the ones you want, or type model names
Prices	USD per million tokens, used only for cost tracking. Fill reference prices fills in common models; set a rate if a reseller discounts official prices
Concurrency	Maximum parallel requests on this channel, unlimited by default. Set 1–2 if you see frequent 429s

You can add several channels — for example a cheap DeepSeek channel for analysis and Claude for the assistant.

Common providers

Provider	Type	Base URL
Anthropic Claude	Anthropic	Leave empty
OpenAI	OpenAI-compatible	https://api.openai.com/v1
DeepSeek	OpenAI-compatible	https://api.deepseek.com
Qwen (Alibaba Cloud)	OpenAI-compatible	https://dashscope.aliyuncs.com/compatible-mode/v1
Moonshot (Kimi)	OpenAI-compatible	https://api.moonshot.cn/v1
Local Ollama	OpenAI-compatible	http://<ollama-host>:11434/v1

Local models

Inside a container, localhost is the container itself. If Ollama runs on the host, use the host’s LAN address, or add extra_hosts: ["host.docker.internal:host-gateway"] to the damail service in Compose and use http://host.docker.internal:11434/v1. The model must produce reliable JSON; very small models categorize poorly.

Pick models

There are three model settings at the bottom of the page:

Setting	Used for	Tip
Default model	Categorizing and summarizing new mail, the AI brief in reports, AI search, label suggestions	Pick something cheap and fast if you get a lot of mail
AI assistant	The chat panel in the corner	Needs tool calling; a stronger model works better
Translation	The Translate button on a message	Can use its own channel and model, with a target language

The assistant and translation follow the default channel and model unless you change them; switching channel means picking a model again.

Effort controls how long the model thinks: “low” is enough for categorizing; higher is slower and uses more tokens. Anthropic supports low / medium / high / xhigh / max; OpenAI-compatible APIs support low / medium / high.

Click Save settings, then turn on Enable AI at the top. The Test button next to each model makes a real call and shows the latency or the error.

Rate limits and retries

• When a provider returns 429, the channel pauses briefly (following Retry-After, 5 seconds to 2 minutes) and then continues with queued requests
• The channel list shows active and queued requests and when it was last rate-limited
• Failed analysis is retried; if it still fails, AI is skipped and inbox rules and notifications still run

Privacy

What is sent to the model

Once enabled, the sender, recipients, subject and body (truncated at 20,000 characters) of each message are sent to the provider you choose. You can limit analysis to some mailboxes, ignore certain senders, or use a local model only.

Every prompt tells the model to treat message content as untrusted data. Model output is only used for labels, summaries and replies that wait for your approval — it never triggers sending. You can read the system prompts for each feature under AI → Prompts.