Agents and models

An agent’s model is any AI SDK v6 LanguageModel, or a gateway model id string. Sendo never picks a provider for you and has no model bundled.

import { openai } from "@ai-sdk/openai";

sendo.agent({
  name: "recepcion",
  model: openai("gpt-5-mini"), // or "openai/gpt-5-mini" through the AI Gateway
  instructions: "Eres la recepcionista de la clínica. Responde breve y en español.",
  pricing: { inputPerMTok: 0.25, outputPerMTok: 2 }, // optional, USD; used for costCents
});

The full definition (tools, handoff, draft mode, limits) is in the API reference.

For exploring: one key, many models#

OpenRouter (through its AI SDK provider) or the Vercel AI Gateway (a string id such as "openai/gpt-5-mini") let you try many models with one key and switch by changing a string. That is the fastest way to find a model that is good and cheap enough for your conversations.

Two costs to know about:

• A markup. These services add a fee on top of the provider’s price. Check their current pricing pages; Sendo doesn’t copy rates.
• An extra network hop. Every turn goes through one more service, which adds latency and one more thing that can fail.

For production: a direct provider#

Once you know the model, call its provider directly (@ai-sdk/openai, @ai-sdk/anthropic, @ai-sdk/google, …): no markup, one hop fewer, and the provider’s own data terms. For sensitive data (health, finance), read those terms and prefer providers with zero data retention agreements.

Changing the model is one line in lib/sendo.ts; nothing else in Sendo depends on it.

What a turn costs#

Every agent turn stores tokensIn and tokensOut on the reply (or draft) message, and costCents:

1. the provider’s reported cost when it gives one (OpenRouter reports it per call);
2. otherwise pricing from the agent definition (USD per million input and output tokens);
3. otherwise null, shown as “—”.

Admins see the cost of each turn in the inbox; Metrics adds cost per day and per agent. See Voice notes and cost.

Keep turns bounded#

• historyLimit (default 20 messages) bounds what the model reads.
• maxStepsPerTurn (default 5) bounds tool calls per turn.
• maxMessagesPerMinutePerContact (default 10) skips turns for floods; messages are still stored.
• To cap output tokens, wrap the model with the AI SDK’s wrapLanguageModel and a middleware that sets maxOutputTokens (the live demo caps every call at 300 this way).

One agent for many organizations#

tenants and accounts restrict where an agent answers, and instructions can be a function of the account’s tenant. See Numbers and tenants.