AI agents

Give your app an AI agent

One address and one key. The agent holds the instructions and data; your app sends the conversation and gets the answer.

Send me the link for my desktop
Create an AI agent

Your app calls one address

The OpenAI Chat Completions format at https://api.light-cloud.com/ai/v1, so most SDKs need only a different base URL. The agent’s Connect tab gives you the key as LC_AI_KEY.

JavaScript
import OpenAI from "openai";

const ai = new OpenAI({ baseURL: process.env.LC_AI_URL, apiKey: process.env.LC_AI_KEY });

const reply = await ai.chat.completions.create(
  { messages: [{ role: "user", content: "Is there a free trial?" }] },
  { headers: { "x-lc-task": "support-reply" } }
);
curl
curl https://api.light-cloud.com/ai/v1/chat/completions \
  -H "Authorization: Bearer $LC_AI_KEY" \
  -H "x-lc-task: support-reply" \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Is there a free trial?"}]}'

What the agent knows

Set on the agent, so a change applies to the next message with no redeploy.

  • InstructionsWho the agent is and how it answers.
  • GuidelinesRules it keeps, whatever it is asked.
  • DataPasted text, web pages and files, up to about 25,000 tokens, sent with every message. The page shows what that costs per message.
  • Try itSend a test message from the console before your app does.

Three ways to make one

  • In the consoleThe AI agent tile on the Create page.
  • From the chat“An agent that answers questions from my price list.” It creates the agent, adds the data and connects your app.
  • From your assistantThe MCP server’s agent tools, from Claude or Claude Code.

Models and vendors

Light Cloud picks the model and vendor per task.

  • Tasks: name each kind of request (x-lc-task: support-reply) and each keeps its own model, usage and history.
  • Keep mode stays on the model and vendor you chose.
  • Save mode moves a task to a cheaper vendor of the same model when that saves 10% or more, and suggests cheaper models for you to approve. A model never changes without your yes.
  • Failover: when a vendor fails before answering, the next one takes the request.
  • Models: Gemini 4, Gemini 3, Gemma 4 and gpt-oss, each with its price on the models page.

Cost

The vendor’s price plus 5%, counted against your plan’s included usage. Your workspace’s one usage limit covers it.

On its own

Add a schedule and Light Cloud calls a URL on your app every 15 or 30 minutes, every hour, every 6 hours or once a day.

Your data

Light Cloud keeps request metadata for 30 days, never prompts or answers. The model providers do not train on these requests without permission; a provider may cache a request in memory for up to 24 hours and log prompts to detect abuse.

Questions

Give your app an AI agent

Create one from the Create page, try it, then connect your app.