# No AI inside

> Checkend has no language model, no model keys and no model calls. You bring your own agent.

Checkend has no AI inside. It has no language model, no model keys and no model calls. It calculates facts. Your agent does the thinking.

## What this means

- Checkend never sends your data to an AI provider.
- Checkend never asks you for a model key, and it has no token bills.
- Checkend never writes text with a model. Each sentence in a report or a brief comes from a fixed template and computed numbers.
- The MCP server does not use MCP sampling. It never asks the model of your client for text.

## Bring your own agent

You connect the agent that you already use. Checkend gives it data and records its work through open interfaces:

| Interface | Example clients |
|---|---|
| MCP | Claude Code, Cursor, any MCP client |
| `ce` CLI | Any coding agent with a terminal, CI jobs, scripts |
| REST API | Your own code, n8n flows, scheduled jobs |
| Webhooks | Any receiver that can start an agent |

See [Connect your agent](/docs/start/connect-your-agent.md).

## Why

| Reason | What it gives you |
|---|---|
| **Privacy** | Data goes to an AI provider only when your own agent asks for it. |
| **No cost and no keys** | You pay no model bills through Checkend, and you manage no model keys in it. |
| **No lock-in** | Use Claude, ChatGPT, a local model, or any client of MCP, the CLI or HTTP. |
| **Trust** | Checkend cannot invent numbers. Each number comes from your data, with its sample size and a test where one applies. |

## Who does what

| Work | Who does it |
|---|---|
| Collect events, sync search data, crawl pages | Checkend |
| Run the rules and make findings | Checkend |
| Measure each change with a statistical test | Checkend |
| Read the brief and decide what to change | Your agent |
| Write the new title, the fix or the pull request | Your agent |
| Record the change | Your agent (an annotation) |
| Approve risky actions | A person |

Checkend keeps the data schedule: syncs, rules, outcomes and webhooks. Your agent keeps the thinking schedule, for example a weekly review. Connect them in two ways:

- **Push:** a webhook starts your agent. See [Webhooks](/docs/guides/webhooks.md).
- **Pull:** your agent runs on its own schedule and calls MCP or the CLI.

## MCP prompts

The MCP server gives three prompts: `weekly_seo_review`, `investigate_traffic_drop` and `launch_check`. A prompt gives your agent the task, the rules of the closed loop and the data. Your model does the work.