> For the complete documentation index, see [llms.txt](https://orbit-pilot.gitbook.io/orbit-pilot-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://orbit-pilot.gitbook.io/orbit-pilot-docs/connect-from-ai-apps/mcp-security-and-guardrails.md).

# MCP security and guardrails

Connecting any AI host to a wallet product introduces real risks. Orbit MCP is designed so the **host can prepare, never sign**. You remain the last gate.

{% hint style="danger" %}
You are responsible for what you approve. Read every Confirm panel. Testnet today; treat every signature as irreversible practice for mainnet habits.
{% endhint %}

## Risks that come with MCP (industry-wide)

| Risk             | Why it matters                                                      |
| ---------------- | ------------------------------------------------------------------- |
| Prompt injection | Malicious text could try to steer the model into harmful tool calls |
| Over-trusted AI  | Users may click Confirm without reading amounts or destinations     |
| Secret leakage   | Bad designs ask the model for seeds, device shares, or signed XDR   |
| Confused deputy  | A connector might try to act beyond what you intended               |
| Phishing UX      | Fake "approve" links outside Orbit's origin                         |

## Guardrails Orbit puts in place

### 1. AI never holds keys

* No device shares, recovery secrets, or raw `S…` secret keys are accepted in MCP tool args
* Tool argument checks reject forbidden fields and secret-looking strings
* Signing happens only in Orbit-origin UI with **your** wallet

### 2. Scopes: read and prepare only

* API / OAuth scopes are sanitized to `read` and `prepare`
* Sign / broadcast style scopes are rejected even if a client asks for them

### 3. You confirm every executable action

* `orbit_do` / prepare creates a **pending action**, not a live submission
* An Orbit Confirm panel (or first-party approve URL) must complete with your signature
* Pending MCP actions expire (about **1 hour**)

### 4. Embed and approve URL protection

* Approve / embed URLs require a short-lived **HMAC `et` token**
* Panels should not load without a valid token
* `frame-ancestors` limits who can iframe Orbit approve surfaces (ChatGPT / Claude domains)

### 5. Same engines as the website

* MCP does not use a separate "shadow" transaction builder
* Intent routing matches Orbit chat so behavior stays consistent and auditable

### 6. Optional wallet binding

* MCP keys can be bound to a specific Stellar `G…` address
* Requests for a different wallet are rejected

### 7. Host instructions

* Orbit tells the model not to paste approve links when the in-chat panel should appear
* Suggested next commands are returned for the model to offer, not to invent amounts

## What you should still do

1. Review asset, amount, destination, and protocol on every Confirm
2. Prefer Orbit passkey / recovery setup you control
3. Revoke MCP / OAuth access if a host account is compromised
4. Use Testnet funds only until you fully understand each flow
5. Never paste seed phrases into ChatGPT, Claude, or Orbit chat

See also: Security disclaimer


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://orbit-pilot.gitbook.io/orbit-pilot-docs/connect-from-ai-apps/mcp-security-and-guardrails.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.