> ## Documentation Index
> Fetch the complete documentation index at: https://docs.neuro-tech.io/llms.txt
> Use this file to discover all available pages before exploring further.

# AI applications

> Connect an AI application to permission-scoped tools on a Neuron

Use Model Context Protocol (MCP) when an AI application needs a small set of callable tools. Use the Agent API when your application code should call Neuro over HTTP.

<Warning>
  The Agent API is Neuro's primary HTTP API; its name is unrelated to AI agents. MCP servers and permissions vary by Neuron, so obtain the endpoint and OAuth settings from its operator.
</Warning>

## Start here

* [Connect Codex](/mcp/quickstart) to one read-only MCP endpoint.
* [Configure a local Neuron](/mcp/local-neuron) when you administer the deployment.
* [Connect to another Neuron](/mcp/remote-neuron) with settings from its operator.
* [Compare MCP servers](/mcp/servers/overview) and choose only the capability your application needs.

## Choose the interface

| Need                                                           | Recommended start                           |
| -------------------------------------------------------------- | ------------------------------------------- |
| A model needs a bounded set of callable tools                  | [MCP](/mcp/overview)                        |
| Application code owns orchestration and a supported SDK exists | [SDKs](/developer-tools/sdks)               |
| Application code needs direct HTTP access                      | [Agent API](/neuron-api/introduction)       |
| The system needs direct XMPP semantics and client-managed keys | [XMPP and email](/protocols/xmpp-and-email) |

## Safety model

* Connect only the MCP server required for the task.
* Grant only the OAuth scopes and Neuron privileges required by its tools.
* Keep approval boundaries around signatures, payments, identity disclosure, and destructive operations.
* Treat MCP safety annotations as planning hints, not authorization.
* Log identifiers and outcomes without logging secrets or unnecessary personal data.

## Identity and payment tools

No documented MCP server currently provides a complete identity-and-payment example. Do not infer those tools from the Agent API. Use only the tools returned by MCP discovery on your Neuron, and require human approval before signatures or payments.
