# Limits, privacy and safety

> What the connector can't do, what the chat won't design, what goes to your AI provider, and how to read Speed's results safely.

Source: https://tryspeed.ai/docs/limits

## Limits

| Limit | What it means for you |
| --- | --- |
| One tool call: 280 seconds | A longer step times out in the chat. The tab may still finish it, and large FE reports can take longer. See [A step timed out](https://tryspeed.ai/docs/troubleshooting.md#a-step-timed-out). |
| Exported files: about 3 MB | Larger files fail through the chat. Open them in the Speed tab instead. |
| Download links: one hour | Ask for the file again after that. |
| One live tab per chat | The tab you opened from the chat's link is the live one; other Speed tabs wait on standby. |
| One tab at a time | A newer Speed tab takes over from an older one. |
| The chat's own model | The work is done by the model on your Claude or ChatGPT plan, with that plan's usage limits. Speed runs no model for connector use. |

## What goes to your AI provider

Your chat's model reads what Speed's tools return, so it goes to your AI provider (for example Anthropic for Claude, or OpenAI for ChatGPT) under your plan's terms:

- **The design's numbers:** part names, sizes, positions, materials and masses, physics check results and FEA results.
- **Pictures of the design** from `view_model`.
- Everything you type in the chat.

The design itself lives in your Speed tab and your Speed account. Speed does not run a model on it for connector use.

> **Controlled data:** Don't use the connector with designs your organisation hasn't cleared for your AI provider: export-controlled, classified or customer-confidential data. Ask your security team first.

## You stay in control

- You watch every step in the Speed tab. Each tool call is one step in the Agent panel.
- **Pause** stops further calls. **Disconnect** ends the connection. **Undo** takes back changes one at a time.
- To revoke access for good, remove Speed from your AI app's connectors.

## What the chat won't design

The AI model in your chat decides what it will help with. Models such as Claude decline to design weapons, or parts whose purpose is to disable or bring down aircraft or harm people. That includes kinetic interceptors, the effectors of jammers, warheads and launchers. Speed does not try to work around this.

Speed is for the rest of the hardware: airframes, sensors, gimbals, optics, structures, mounts, power and enclosures. For counter-drone work, design **detection and tracking**: masts, antennas, optical sensors, their power and their structure. See [the detection mast example](https://tryspeed.ai/docs/examples/drone-detection-mast.md).

If the chat declines part of a request, it usually offers the part it can do. Take that, or rephrase your request around sensing and structure. See [Troubleshooting](https://tryspeed.ai/docs/troubleshooting.md#the-chat-declined-part-of-my-request).

## How to read the results

- Speed's physics checks and FEA are **screening grade**. They are quick estimates for choosing between concepts, on simplified part shapes.
- An engineer must review a design, and its loads and supports, before anyone builds it.
- Never describe a Speed design as verified, certified or flight-ready. The chat is told the same.
- Catalogue parts carry datasheet figures. Check them against the maker's current datasheet.
- Defence hardware is regulated. Check export rules and local law before you share, build or fly a design.