This website
This website is made of static pages. The hosting and the CDN in front of it process IP addresses and access logs for every request, as any web server does.
A visit counter from statinside.com records page visits, so we can see which pages are read. It counts visits to this website only and plays no part in the app.
Pressing the download button records the time, your IP address, your browser's user agent and the referral source, if there is one, so downloads can be counted. After that the website redirects you to the installer. Reading the pages without pressing the button creates no download record.
The FlowLatch app
FlowLatch runs locally on your PC and sends no telemetry. Your chats, your working folders and your API keys stay on the PC, and no account is needed to run the app. Every action the agent takes appears in the chat as a card, so you can see what was run, and an expanded card shows the full output or the diff.
Secrets that appear in tool output are redacted before the output reaches the model. Prompts go only to the model endpoint you pick, so the privacy of your project depends on that choice. Dangerous commands are blocked.
With a cloud model, the prompt, including the text of any file the agent reads, goes to that provider and is handled under the provider's terms. With a local model nothing leaves the machine. The built-in model needs no endpoint of your own.
Optional features that use the network
Web search lets the agent look something up, for example a page of your test framework's documentation, and it can be routed over an anonymizing network if you choose that. http_get downloads one page when the agent needs to read it.
Control of your chats from a phone browser uses one outbound connection from the PC, and you can switch it off. Update checks also use the network, so the app can show the Update button when a new version is ready.
A cloud AI provider is a network feature as well, because every prompt travels to its endpoint. Shell, web and SSH are not limited to the working folder, so keep the work to systems you own or are authorised to use.
Your choices
You choose the model, and with it where your prompts go: a cloud provider, a local model on your PC or the built-in model. If a project must not leave the machine, use a local model and keep real customer data out of the working folder anyway.
You choose the approval mode. Auto-run acts at once, Ask first pauses before actions that change something and shows the diff, and Plan only writes a plan and runs nothing. A stricter mode gives you more chances to stop an action before it happens.
You can switch off the phone-browser control and leave web search unused. Tests should use test accounts and fake data, so that no real person's information is ever part of a prompt. The app asks for no account, so there is no profile that could be linked to your chats, and the only identifiers this website sees are the technical ones listed above, such as the IP address and the user agent.
Contact
To report a problem, ask a question or tell us about something unclear in these notes, use the Send feedback button inside the app. That is the one contact channel for FlowLatch.
This page describes version 2.9.30 of the app and the website as of 8 November 2026.
Updated 8 November 2026
