Web

Web chat is the richest way to work with your agent. Watch it work in real time, hand it files, and see everything it produces. On a Mac, the desktop app is this same interface plus access to your local files and apps.

​

Mobile

Skydive works as a progressive web app on iPhone. Open skydive.com in Safari, tap Share, then Add to Home Screen. The app installs without going through the App Store, launches full-screen, and stays in sync with your other sessions. A native iOS app is planned for the future.

​

What you get

​

Sending attachments

Use the composer to attach files. The agent downloads and works with them (reading a contract, transcribing audio, analyzing a spreadsheet) and returns results you can download from the chat.

​

Media in replies

When an agent shares images, GIFs, or videos, they render inline as a gallery rather than as download cards. Up to ten items appear in a responsive grid; a +N more label collapses anything beyond that. Click any tile to open it full-size. If you want a downloadable file card instead, the agent can pass --download to platform web attachments share.

​

Diagrams

Mermaid diagrams render inline in web chat. When an agent responds with a fenced code block tagged ```mermaid, the chart draws directly in the message rather than showing raw text. Flowcharts, sequence diagrams, entity-relationship diagrams, and every other Mermaid diagram type are all supported. The same rendering applies to messages you send, assistant replies, and the transcript viewer.

​

Conversations

Each chat is its own thread. All threads share the agent’s memory. Start a new conversation for a new topic; the agent still carries everything it has learned.

​

Archive and delete

Archive a finished thread to clear it from your main list without losing it. Archived chats move to an Archived section in the sidebar and become read-only. Restore one at any time to pick up where you left off. Archiving is private to you: in a shared chat, everyone else still sees it as active. Deleting is permanent, so Skydive asks you to confirm first. In a shared conversation, you can only delete it when you are the last person in it.

​

Context compaction

When a conversation runs long, the agent automatically compacts earlier context to stay within its working window. Nothing is lost: the agent carries a summary of what was compacted forward, and the full transcript stays in your history. Web chat surfaces compaction inline. A ✓ Compacted this conversation marker appears in the timeline at the point where compaction happened, with a timestamp. While compaction is in progress you’ll see a live indicator. No action is needed on your part.

​

Fork a chat

Fork a chat to copy the whole thread into a new conversation and branch off without touching the original. The copy carries the full history up to that point, so the agent keeps all the context, and you can take it in a different direction while the source stays as it was. Find it in the conversation options menu (the dots next to the title), alongside Rename chat. A fork also frees a thread from its channel. Forking a Slack conversation gives you a copy you can carry on in web chat, which is handy when you want to keep working from the browser on something that started in Slack.

​

More than one agent in a chat

A web conversation can hold up to three agents. When work needs a specialist or a second opinion, an agent can pull a peer into the same thread. The peer sees the conversation and takes part directly, alongside you. Added and removed agents are announced in the thread, so it is always clear who is present. You can also ask an agent to bring in or drop another agent. This applies to web chat only; Slack, email, and text have their own membership.

​

Switching channels

When you are chatting in web and want to continue on iMessage, Slack, or email, look for the channel pill at the top of the conversation. Connected channels show an Open in [channel] action that takes you straight into that thread.

​

When a run fails

Most of the time a failed run is self-explanatory: the agent tells you, you retry, it works. Two kinds of failure are worth understanding. Usage and limits. If your workspace is out of credits or over a rate limit, the error says so and links to usage and billing. Top up or wait for the limit window to reset. Model provider errors. Sometimes the model behind the agent is what fails, and the error names the fix when there is one. The one to know today: Claude subscription connections refuse requests until model training is verified off. The fix is in Claude, not Skydive: open Claude’s Settings > Privacy and turn off Help improve our AI models, then retry. The error copy shows the same steps. Where the message appears depends on where you’re talking to the agent: live in the conversation as it fails, in run history afterwards, and in a failure notice on iMessage threads. The copy is the same everywhere, so a fix you look up once applies anywhere the error catches you. Providers add new failure modes over time. When the error copy already names a fix, follow it. When it’s the generic “model stopped responding” line, retry once; if it keeps failing, the problem is the provider connection, not your prompt.

​

When to use web chat

Use web chat when the work is substantial: building something, processing files, a long back-and-forth, or anything where watching the agent work helps. For quick asks, Slack or text may be faster.