The chat UI isn't the product

I've been tracking the "AI-native X" launches on X for about a year.

AI-native Notion. AI-native Linear. AI-native Figma. AI-native CRM. AI-native spreadsheet. AI-native inbox. Every one of them eventually shows me the same screenshot: a left sidebar with a file tree, a main canvas, and a chat panel on the right.

The chat panel is where "the AI lives." You type what you want. The agent does something. You wait.

I think that bet is going to lose. This post is what we built instead.

The interface war is already over

Your user already has an agent they love. Maybe it's Claude Code in the terminal, Cursor running in the background, Codex, Windsurf, a local Ollama setup with a homegrown wrapper. Whichever one it is, they picked it deliberately and they are extremely loyal to it. You can tell because they will argue about it in every discord you're in.

If you're launching "AI-native [category]" in 2026, your chat panel has to beat the one they already have. Your keyboard shortcuts, your context handling, your model routing, your multi-step planning. All of it, better than what Anthropic and OpenAI and Cursor ship full-time with three-figure headcounts.

You are not going to win that fight. I am not going to win that fight. Nobody outside a handful of well-funded labs is going to win that fight.

What you can win

What none of those agents have is hands.

Claude Code is phenomenal at code. It is blind to your team's shared sheet, your company mailbox, your slide deck, your tasks, your files. If it decides the right next step is "send the customer a one-paragraph update and log the interaction," it cannot do that. It has to write you the paragraph and wait for you to click through three apps.

The other thing none of those agents have is a place for their output to land. When your agent produces 40 artifacts in a morning (drafts, rows, replies, notes), they scatter. Some into your inbox. Some as CSV blobs in a chat. Some as a diff you have to apply by hand. No shared surface. No diff view. No approval step. No record of what changed when three teammates' agents all worked on overlapping stuff overnight.

So we built those two things. Neither of them is another chat UI.

SpatioMCP, the hands

SpatioMCP is an MCP server. It exposes nine primitives (notes, sheets, slides, mail, tasks, calendar, files, chat, contacts) as tools your agent can call. The config is four lines:

{
  "mcpServers": {
    "spatio": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.spatio.app/mcp"]
    }
  }
}

Drop it into Claude Code, Cursor, Codex, whichever client you're in, and your agent is suddenly capable of creating notes, updating cells in sheets, drafting replies in your inbox, blocking time on your calendar. All from the same chat window it was already in.

This is the half most people expect. It's also the half getting the most imitators. Every other week there's a new "Spatio-like MCP for X." Good. The category should exist.

Spatio Desktop, the workbench

This is the half nobody's shipping, and honestly the half I'm more excited about.

Desktop is where every artifact your agent creates lands. A shared feed for your team. You open it and you see: the three emails your agent drafted this morning, the 14 rows it added to the forecast, the note it wrote summarizing last week's tickets. You review. You edit. You approve. You send.

It is deliberately not a chat interface. Your team does not prompt agents in Desktop. That is what Claude Code and Cursor are for. Desktop is downstream of the prompt. It is the collaborative review surface where what your agent produced becomes what your team actually ships.

It looks like Gmail if Gmail were for agent output. That's roughly right.

The reason Desktop exists: at team scale, the bottleneck is not draft quality. The bottleneck is review throughput. If every agent output lands in a different app, nobody can triage it. If every teammate's agent works independently, nobody sees what changed. Desktop fixes both. One surface. Diffs. Accept, edit, reject. Reversible.

The bet

The interface your user already loves will eat whatever interface you build. So don't build another one. Build what that interface is missing.

What it's missing is hands (your agent can't actually do anything outside its chat window) and a workbench (your team can't keep up with what the agent produces).

We ship both.

You can drop SpatioMCP into Claude Code in about 90 seconds. Sign up and Desktop opens with an empty feed that fills as your agent starts working.

Use it for a week. You'll see what the product is actually about.

Matt