Stored agents

A stored agent keeps its model, instructions, tools and widgets setting in Cencori. Anything that runs it, by id, gets the same agent: change it once and every app using it changes.

POST /v1/agents
CENCORI_API_KEY: csk_...
 
{
  "name": "Sales analyst",
  "config": {
    "model": "gpt-oss-120b",
    "system_prompt": "You are a concise sales analyst. Answer with charts and tables when it helps.",
    "tools": ["web_search_preview"],
    "widgets": true,
    "temperature": 0.3
  }
}

{ "id": "fea49590-...", "name": "Sales analyst", "is_active": true, "config": { "model": "gpt-oss-120b", "tools": ["web_search_preview"], "widgets": true, ... } }

config.tools holds built-in tools (web_search_preview and others). Your own function tools stay in your code: add them when you run the agent from the SDK.

From the SDK, with your own tools on top:

import { Agent, run } from 'cencori/agents';
 
const analyst = Agent.fromCencori('fea49590-...', { tools: [getSales] });
const result = await run(analyst, 'How is revenue trending?');

Over HTTP, start the session with the agent:

POST /v1/sessions
 
{ "agent_id": "fea49590-...", "turn": { "input": "How is revenue trending?" } }

The stored model and instructions apply to every turn in the session; a turn's own model or instructions only fill in what the agent doesn't set.

From a browser, with no backend: see below.

A publishable key (cpk_) can run a stored agent from a web page. Create one in API keys and set its allowed domains: requests from any other site are refused.

import { AgentChat } from 'cencori/react';
 
<AgentChat agent="fea49590-..." publishableKey="cpk_..." greeting="Ask about sales" />

What a publishable key can do is deliberately small:

• Start a session only with a stored agent (agent_id is required), and continue only sessions it started.
• Send the user's message. The agent's own model, instructions and tools always apply; anything else in the request is ignored.
• Conversations are capped like the website widget's: a turn limit per conversation and a pace limit per minute.
• It can't list sessions, read other sessions, approve tool calls, or attach tenants or users.

Usage is billed to your project, like any other request. Use a backend (Agent UI server) when the agent needs your own tools, approvals, or to know who the user is.