Nine new AI sidebars, one chat that reaches all of them

At least nine. Notion shipped voice input on desktop (v7.10.0, April 6), shareable AI chats (April 7), database tab customization (April 10), and Notion 3.4 Part 2 with Workers for Agents and the Views API on April 14. Todoist shipped AI-powered views. Linear expanded sub-issue cascade automation. Fantastical added AI scheduling suggestions. Microsoft 365 launched Power Apps 2026 wave 1 and a Current Channel update on April 10. Apple shipped a major Logic Pro upgrade and a major Pixelmator Pro upgrade on April 9. Obsidian, Coda, and ClickUp pushed meaningful changes too. Most of these added some form of new AI control in the desktop UI, which is the entire reason this guide exists.

Because the pattern shifted. Through 2025, productivity apps mostly added an AI text box on the side. In April 2026 they started embedding AI as infrastructure: Notion's Workers for Agents, Linear's cascade automation, Fantastical's scheduling reasoner, Logic Pro's audio reasoning. Each one ships its own surface, its own button placement, its own opt-in flow. The result for users is fragmentation. You either learn nine new sidebars or you find one tool that talks to all of them. This guide is about the second path.

Every native and Electron app on macOS publishes its UI controls into a structured tree called the AX tree, accessed through AXUIElement APIs. When Notion ships a new mic button in v7.10.0, the button enters the AX tree the moment the build is installed. When Logic Pro adds a new mixer toggle, the toggle enters the AX tree the moment the user opens the new release. Reading the AX tree is faster than vision, more accurate than OCR, and does not require any vendor API. It is what makes a single agent able to drive nine different productivity apps without nine different integrations.

Through a bundled binary called mcp-server-macos-use. It is shipped inside the Fazm.app at Contents/MacOS/mcp-server-macos-use and registered as an MCP server in nine lines of acp-bridge/src/index.ts (lines 938 to 946). It identifies as SwiftMacOSServerDirect 1.6.0. When the agent loop queries tools/list, the binary returns exactly six tools every time: open_application_and_traverse, click_and_traverse, type_and_traverse, press_key_and_traverse, scroll_and_traverse, refresh_traversal. Each one operates on whatever element appears in the AX tree of the focused app. Add a new control in any of the nine apps and the same six verbs reach it on day one.

Two places. First, acp-bridge/src/index.ts:938-946 is the registration block that mounts mcp-server-macos-use as an MCP server with no args, no env, and no per-app code. Second, Desktop/Sources/AppState.swift:439 is where the Swift app reads the focused window of whatever productivity app is in the foreground via AXUIElementCreateApplication(frontApp.processIdentifier) and then AXUIElementCopyAttributeValue with kAXFocusedWindowAttribute. That call returns Notion's window when Notion is focused, Logic Pro's window when Logic Pro is focused, Fantastical's window when Fantastical is focused, with no per-app branch.

No. That is the entire point of using accessibility APIs instead of vendor integrations. Notion's voice input button has no public REST endpoint. Fantastical's AI scheduling suggestion has no public REST endpoint. Logic Pro's new mixer features have no public REST endpoint. None of those facts matter when the agent is reading the AX tree. The button is a node, the node has a role and a label, and click_and_traverse can target it by element=label without any vendor cooperation.

A screenshot agent captures a 1 to 2 MB PNG of the screen and asks a vision model to guess where to click. The model can mistake icons, hallucinate buttons that are not on screen, or get confused by overlapping windows. Fazm reads the AX tree directly. Each line in the tree is in the format: [Role] "text" x:N y:N w:W h:H visible. The role is one of AXButton, AXTextField, AXMenuItem, AXTab, AXCheckBox, etc. The text is the accessibility label. The coordinates are exact. There is no inference, no model cost, no flicker. New buttons that landed in the April 2026 productivity wave appear in the tree the same as any other button.

No. Fazm itself shipped nine numbered releases between April 3 and April 16, 2026, but none of them were vendor adapters for Notion, Todoist, Logic Pro, or anyone else. Those changes were chat UX, pop-out window plumbing, the ACP v0.25.0 protocol upgrade, and a Smart/Fast model toggle. The accessibility surface that reaches the new productivity buttons is unchanged because it does not need per-vendor code. Six verbs, an AX tree, no adapters.

Yes, with high confidence. The number six is structural. Each verb maps to one fundamental action on a generic GUI: open an app, click, type, press a key, scroll, re-read the tree. New productivity features in May 2026 will change the labels and roles on AX nodes. They will not change which verbs the agent has to perform. The same was true through Q1 2026, through April 2026 (nine app updates, zero binary updates), and there is no pending macOS API change that would invalidate it.

Yes. Once Fazm is installed, run ls -la /Applications/Fazm.app/Contents/MacOS/mcp-server-macos-use. Or run the binary in stdio mode and send {"jsonrpc":"2.0","id":1,"method":"tools/list"} to it; it returns the same six-tool list described above. The version line says SwiftMacOSServerDirect 1.6.0 as of this writing.