ShogunAI

ShogunAI is a macOS-native personal AI layer that remembers your workday and acts on it. Built for Apple Silicon, it runs on-device, captures context from your apps through the macOS accessibility layer, and executes tasks inside the tools you already use — all without sending your raw data to a server. The product is currently in early-access waitlist stage, with the site reporting over 570 people have joined.

What It Is

ShogunAI positions itself as a "personal AGI on your PC" — not a smarter model, but a memory and execution harness layered around the frontier models you already use. It solves two problems the team identifies as persistent gaps: the cold-start problem (every AI session requires re-explaining your context) and the answer-not-outcome problem (AI gives you a draft, but you still do the filing, sending, and follow-through). ShogunAI addresses both by maintaining a live, private memory of your work and acting on it through a structured execution layer with explicit approval gates.

How the Memory Layer Works

ShogunAI reads text from the macOS accessibility layer — the same channel screen readers use — and indexes it on-device using a bundled multilingual embedding model. Key mechanics:

• Tiered memory: the last day is kept in full for fast recall; the last month in full detail; older content is compressed
• Visual recall: an optional compressed screenshot is taken only where text capture returns nothing; off by default, encrypted, and auto-deleted at a retention age you set (1–7 days or a custom duration)
• Meeting audio: streamed to a speech service for live transcription only; never written to disk; only the transcript persists
• Source and time attached: every recalled item carries where it came from and when, so answers can be verified rather than trusted

The Execution Layer and Autonomy Model

The execution layer operates under three named levels of autonomy: reversible work (runs automatically), drafts (wait for user review), and anything that leaves the Mac (requires explicit approval before sending). The site states that "no prompt moves that line" — the approval boundary for outbound actions is fixed by architecture, not by configuration. Every action leaves a traceability log recording what ran, on what evidence, and what left the device. Third-party hops are marked in the log.

Supported execution actions include inline drafting (press Option to write a continuation at the cursor in any app), meeting minutes with proposed next actions, two-way live translation, daily briefs, and cross-tool recall over mail, calendar, chat, and docs.

Privacy and Deployment Architecture

ShogunAI's architecture is explicitly local-first:

• Capture, indexing, and search all run on the Mac with no server-side copy
• Encrypted local store; no training on user data
• BYOK (bring your own API key) supported — keys live in the system keychain
• Users can also run on an existing assistant subscription (e.g., ChatGPT, Claude) without a separate API key
• What does leave the device is scoped to the specific slice needed for a request, logged with its purpose, and marked when it passes through a third party

The FAQ confirms the product is macOS-only and built on native macOS capabilities — the accessibility layer, the notch panel, and the system keychain — running on Apple Silicon. The team states a version for another platform "would be a different product, not a port."

Current Status and Recognition

ShogunAI is in early access as of 2026, with a waitlist open at no cost. The site notes it won the YC RFS Hackathon 2026 (presented by Transpose) and is a member of the NVIDIA Inception Program. It is also featured on Product Hunt. The team describes itself as "a small product and engineering team focused on privacy-first personal AI." Integrations shown include Gmail, Notion, Google Calendar, Slack, GitHub, Linear, Figma, Stripe, Salesforce, Perplexity, Discord, Asana, Dropbox, Vercel, and Claude/ChatGPT/Grok as model providers.