Yansu

Overview

Yansu, by Isoform, is a desktop AI agent that learns by watching. Rather than asking a user to map out every workflow upfront, the app sits in the background on Mac, Windows, or Linux and learns from screenshots and/or supported messaging apps while the user works normally. When it spots a pattern that repeats, it offers to turn that pattern into an automation, a scheduled job, or a small bespoke app — generated and runnable on the same machine.

The framing the team uses is "AI that learns how you work and turns it into software." The product sits in the broader AI workflow automation category but takes an unusual position on it: instead of giving a builder a canvas (n8n, Make, Zapier), it gives a worker an observer that authors the canvas for them, then runs the result.

A second design choice is local-first. The model runs against on-device OCR and audio processing, redaction of sensitive data happens locally, and the published privacy posture is "Nothing leaves your device without explicit permission." That stance is backed by SOC 2 Type II and ISO 27001 certifications — official trust signals for privacy-sensitive desktop AI evaluation.

Key Features

• Observation-driven workflow capture — Watches files, messages, and apps to identify repeated patterns. Surfaces automation candidates instead of asking the user to write a workflow definition.
• Hand-Off automation — Once a pattern is confirmed, Yansu can execute it autonomously using a virtual cursor that never touches the user's real cursor, so foreground work isn't interrupted. The free tier includes ~2 Hand-Offs per month; paid tiers scale up.
• Bespoke desktop app generation — Generates small custom apps and dashboards from observed patterns (spend trackers, meeting recap utilities, chat dashboards). Lives alongside Hand-Off as a way to materialize tribal knowledge into UI.
• Background Computer Use — Runs in a parallel context on a virtual cursor rather than hijacking the screen, which is meaningfully different from agents like Anthropic Computer Use or OpenAI Operator that take over the active session.
• Cross-platform support — Available for macOS 13+ (Apple Silicon and Intel), Windows 10+, and Linux (Ubuntu 20.04+, x64 and ARM64). No mobile, intentionally — the surface is the workstation.
• On-device OCR, audio, and redaction — Sensitive data is redacted locally before any cloud call; the published material claims "nothing leaves your device without explicit permission." SOC 2 Type II and ISO 27001 back the certification posture.

How to Get Started

1. Download for your OS — macOS 13+, Windows 10+, or Linux (Ubuntu 20.04+); direct download, no waitlist.
2. Sign in and grant observation permissions — Yansu may require OS-level screen/automation permissions during setup; describe this generally and avoid comparing permission scope to other apps unless verified.
3. Work normally for a few days — The app observes and accumulates a memory ("tribal knowledge") of how the user actually works. Nothing acts autonomously yet.
4. Confirm the first proposed Hand-Off — When Yansu surfaces a repeated pattern, the user reviews and approves before the agent runs it. The virtual-cursor execution starts here.
5. Connect chat or data sources for richer context — Slack, Microsoft Teams, Feishu, Discord, Telegram, WhatsApp, and WeChat are listed as integrations. Jira, Notion, and Sheets appear as examples of data sources for generated project trackers; verify connector depth before promising them as current integrations.

How It Compares

• vs. Anthropic Computer Use / OpenAI Operator — Yansu's officially stated difference is that it uses its own virtual cursor/background computer-use layer; avoid asserting current Anthropic/OpenAI cursor behavior unless a source is added. That positioning makes it less general-purpose for novel tasks but less disruptive for repeats.
• vs. Manus / Simular / autonomous agents — Those handle a one-shot autonomous task end to end. Yansu's value lives in the learning loop — it's optimized for "tasks I do every week" rather than "a complex task I want done once."
• vs. n8n / Make / Zapier — Those are builder canvases the user authors. Yansu authors the workflow from observation; the trade-off is less explicit control but no upfront workflow modeling.
• vs. Notion AI / built-in app automation — Built-in automations are scoped to one product. Yansu spans whatever the user actually uses across the desktop, which is the typical real-world workflow shape.

The honest read: Yansu is best when there's clear weekly repetition the user can't easily map out. It's a poor fit when the workflow is novel each time or already explicitly modeled in a builder tool.

Pricing & Plans

The official pricing uses approximate monthly allowances for Hand-Offs, crystals, and memories/knowledge; avoid saying they are not strict caps unless the vendor confirms how overages work. The free tier is usable for evaluation but burns through quickly once Hand-Offs cover real recurring work.

Best For

• Operators, ops leaders, and knowledge workers who watch the same weekly cycle of files/messages/apps and want it automated without writing workflows
• Small businesses and indie founders running multi-tool stacks (Notion + Slack + Sheets + Jira) where bespoke internal tools would normally require an engineer
• Power users of AI productivity tools who've outgrown text-based prompts and want an agent that acts on the actual desktop
• Privacy-sensitive professionals comfortable granting OS-level observation in exchange for on-device redaction and SOC 2 / ISO 27001 posture
• Teams comparing local-control alternatives to Manus My Computer before standardizing on a desktop AI agent

FAQ