A business process automation tool that picks the process to automate by watching 60 minutes of your screen.

Those tools assume the user already knows which process to automate. The user (or a process architect) opens a workflow designer, drags BPMN nodes, wires actions, maps fields, publishes the workflow, then runs it on the platform. Fazm inverts the order. The screen observer records 60 minutes of active-window activity at 2 FPS (SessionRecordingManager.swift line 77), uploads those chunks to Gemini Pro at the one-hour mark (GeminiAnalysisService.swift line 213), and the prompt instructs the model to pick the single most impactful task that is currently eating the user's time (line 13 of the same file). The task is saved to the local observer_activity SQLite table (AppDatabase.swift line 960) and surfaced in the Discovered Tasks tab. The user clicks 'Run,' not 'Build.' The artifact is a sentence, not a .bpmn file.

No. There is no canvas, no nodes to wire, no gateways to model. The 'workflow' is a single-sentence task that the agent decodes at runtime. The agent decides per step which of the five bundled MCP servers (fazm_tools, playwright, macos-use, whatsapp, google-workspace; spawned at acp-bridge/src/index.ts line 2550) should handle the step. Parallel steps are just sequential tool calls the agent batches. There is no persistent workflow file to version or deploy. If you need a versioned workflow, Fazm is the wrong tool. If you want the thing done, Fazm is the shape built around that.

It means the bundled macos-use MCP server (acp-bridge/src/index.ts line 1057 spawns mcp-server-macos-use from Contents/MacOS) reads AXUIElement trees from the macOS Accessibility APIs, the same APIs VoiceOver uses. That gives the agent programmatic access to every button, text field, and role in every native Mac app, including Finder, Mail, Notes, Xcode, Figma, Slack, Notion, Linear, and any Catalyst or SwiftUI or AppKit app. The browser is one MCP server among five, not the whole product. A BPA task that touches a Finder rename, then a Slack DM, then a Stripe invoice download, then a Google Sheet update lands as a single user turn chaining all four servers in order.

Three concrete differences. First, the agent reads structured data: role, label, value, children, with a stable ref token per element, so it never has to guess pixel coordinates. Second, it survives app redesigns that would break OCR templates and vision models that memorized pixel regions. Third, it is fast. An AXUIElement tree is microseconds to read; a screenshot pass through a vision model is hundreds of milliseconds per step. For a bookkeeping automation with 40 clicks, that difference compounds into minutes per run.

Line 13 of /Users/matthewdi/fazm/Desktop/Sources/GeminiAnalysisService.swift reads verbatim: 'You are watching ~60 minutes of a user's screen recording. Each video clip captures the active window of whatever app the user was using at that moment. Your job is to identify the ONE most impactful task an AI agent could take off their plate.' The word ONE is load-bearing. One task per analysis, not a dashboard of ten. The model is then forced at line 30 to run two mandatory SQL queries (observer_activity and chat_messages) before proposing anything, and the prompt at line 34 instructs the model to 'err on the side of NO_TASK when in doubt about similarity' to prior suggestions.

The observer runs while the app is open and the screen observer setting is on. handleChunk (GeminiAnalysisService.swift line 166) short-circuits if shortcut_screenObserverEnabled is false (line 168). Chunks are 60 seconds of H.265 video of the currently focused window, written to ~/Library/Caches/observer-recordings, and copied into ~/Library/Application Support/Fazm/gemini-analysis/chunks on finalization. They stay on disk until the 60-minute buffer threshold triggers analysis (line 213), at which point they are uploaded to the Gemini File API, the model picks a task, and the successfully-analyzed chunks are deleted (line 282). A failed analysis keeps the buffer intact and imposes a 5-minute retryCooldown (line 78).

Because suggestion spam kills adoption. The prompt at line 38 says literally 'A wrong suggestion is worse than no suggestion.' Line 42 adds: 'Only flag a task if ALL of these are true' followed by six conditions, including 'you can clearly see what the user is doing' and 'the task is NOT already being handled by Fazm's agent'. A model that returns UNCLEAR fifty percent of the time is more useful than one that always flags something, because the one it does flag is rarely garbage. This is a product decision encoded in the prompt, not in retrieval or post-filtering.

A row with a wand-and-stars icon, the one-sentence task title, a status badge (pending, shown, acted, dismissed), and an expand caret. Expanding reveals the description (3 to 5 sentences), the markdown document write-up with 'What Was Observed', 'The Task', 'Why AI Can Help', and 'Recommended Approach' sections (GeminiAnalysisService.swift line 61), and a 'Run' button. DiscoveredTasksSection.swift polls observer_activity every 10 seconds (line 11) so a newly-analyzed task appears without a manual refresh. Status transitions from pending to shown the moment the row is expanded (markAsRead, line 88).

Process mining tools extract process event logs from ERP systems (SAP, Oracle, Salesforce) and reconstruct the canonical path through the process from timestamps and transaction IDs. Fazm does something different and smaller: it looks at your screen, not your database, and identifies one candidate task that the AI could take off your plate, not the whole topology of a process. Celonis tells an enterprise 'your procure-to-pay has 14 deviation paths, here are the bottlenecks.' Fazm tells one person 'you just spent 40 minutes copying invoice numbers from Stripe into a Google Sheet; want me to do that?' Different altitude, different customer, different price tag. Complementary, not competing.

Yes. The discovered-tasks pipeline is one entry point into the agent; the floating bar is the other. Typing 'open Stripe, pull the last three invoices, rename by date, drop in Dropbox' into the floating bar kicks off the same agent turn the observer would have, except the user authored the task instead of Gemini. The observer exists so that users who would not have thought to ask still get a suggestion. It is an onboarding slope, not a gate.

It persists. observer_activity rows live in ~/Library/Application Support/Fazm/fazm.db, which is not cleared on quit. The buffered chunks that have not yet been analyzed are also persisted, to the buffer-index.json file at ~/Library/Application Support/Fazm/gemini-analysis/buffer-index.json (GeminiAnalysisService.swift line 138). On relaunch, restored entries that no longer have a file on disk are pruned (line 146). Orphaned .mp4 files are cleaned up (line 156). Buffer index survives a crash or a reboot, so a 45-minute recording does not become a 45-minute loss.

Five, spawned at app launch by acp-bridge/src/index.ts and logged on line 2550: fazm_tools (the Fazm HTTP tools for file indexing, user profile, memory graph), playwright (the Playwright MCP with --extension, attached to your real Chrome via the Chrome Web Store extension), macos-use (AXUIElement-based native Mac app control), whatsapp (the bundled WhatsApp MCP binary), and google-workspace (the bundled Python MCP for Gmail, Calendar, Drive, Docs, Sheets). Any custom server declared in ~/.fazm/mcp-servers.json is also loaded on startup (line 1104). The agent picks which server per step, so one user turn can chain several.

Consumer-shaped: signed, notarized Mac app downloaded from fazm.ai, one-click install, no admin role, no CoE. Enterprise pilots run through cal.com/matt364/fazm-enterprise-demo where we walk through the macos-use MCP, the observer's local-only posture, and what an enterprise deployment would look like for a team where each person runs Fazm on their own Mac but shares MCP server configs through ~/.fazm/mcp-servers.json. The architecture is per-user, so adding seats is copy-paste; the workflow-admin role common in UiPath / Appian deployments does not exist in Fazm.