Yuzu: a local AI agent for your Mac

Two rooms, one cat

Cowork for the day. Code for the repo.

Switch between a quiet writing room and a hands-on workspace. Both share the same model, the same harness, and the same rules.

Cowork

Everyday work across your tools: a morning Slack digest, inbox triage, release notes, research briefs. Every task runs in a private sandbox folder, and any task can run on a schedule.

💬 Slack digests🗓️ Scheduled tasks🔎 Research

Code

Open any folder and work on it. Yuzu reads, searches, edits, and runs commands there, pulls context from GitHub, Sentry, or Linear, and asks before it changes anything.

📁 Any folder🧪 Shell and tests🐛 Crash to fix

MCP and skills

Plug in your stack. Teach it your playbook.

MCP connects Yuzu to the tools you already use. Skills teach it how your team uses them. Together, a model on your Mac does real work across Slack, GitHub, Notion, and more.

💬 Slack

“Summarize what I missed in #eng since yesterday and draft replies to anything that mentions me.”

🐙 GitHub

“Review the open PRs, check each against our style skill, and leave a summary on every one.”

📐 Linear

“Turn the TODOs in this folder into issues, tagged by area and linked to the right files.”

🚨 Sentry

“Pull last night's top crash, find the code path in this repo, and propose a fix.”

📝 Notion

“Write this week's release notes from merged PRs and file them in our changelog.”

🌐 Web search

“Compare three alternatives to our job queue in a table and save it next to the design doc.”

Connect in one sentence

Ask “install the DuckDuckGo MCP server” and the agent writes the entry, connects it, and starts using it. Token-based servers go in Connectors or mcp.json:

"slack": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-slack"],
  "env": { "SLACK_BOT_TOKEN": "xoxb-…" }
},
"github": {
  "url": "https://api.githubcopilot.com/mcp/",
  "headers": { "Authorization": "Bearer …" }
}

Local (stdio) and remote (HTTP) servers
OAuth sign-in with PKCE, tokens in the Keychain
Server prompts become /commands
Every MCP call asks before it runs

Teach it a workflow once

A skill is a folder with a SKILL.md and any scripts it needs. Commit it to .yuzu/skills/ and your whole team gets it.

---
name: standup
description: Use for a standup or daily update.
slash: true
---

1. Read my commits since the last workday.
2. Read Linear issues assigned to me that moved.
3. Read my messages in #eng and #product.
4. Write Done, Doing, Blocked. Link everything.
5. Post to #standup after I approve.

Loaded only when a request matches
Run directly as /standup
Pair with Scheduled for a 9 AM digest
Built-in customize-yuzu edits Yuzu itself

What's inside

A real harness, not a chat box.

A complete agent loop with tools, permissions, memory, and metrics, built natively in Swift for Apple Silicon.

Fully on-device

Runs on uzu. Yuzu picks the largest Mirai-quantized model that fits in half your RAM and downloads it once.

Live turns

Thinking, text, and tool calls stream in as they happen. Replies render as Markdown with themed code blocks.

Honest metrics

Context in use, decode tokens per second, and the CPU and RAM used by the model process itself.

Any MCP server

Local and remote servers, OAuth sign-in, Keychain tokens. Tools, prompts, and resources all reach the agent.

Skills and subagents

Team workflows as SKILL.md files with their own scripts, run as /commands. Hand side quests to a subagent.

Nothing gets lost

Every message lands in SQLite before the next turn. Long threads compact into a summary and keep the recent tail.

You hold the keys

Three permission modes. One click.

Cycle with the button under the Code composer. The Plan agent stays read-only in every mode.

CarefulWorkspace ReadReads freely. Asks before every edit and every shell command.

DefaultWorkspace WriteEdits files in the workspace. Still asks before running commands.

FearlessYOLONo prompts at all. For sandboxes, throwaway repos, and brave afternoons.

Get started

Two minutes to a local model.

Download the DMG and drag Yuzu into Applications.
Open it once. This build is ad-hoc signed, so macOS asks first. Open System Settings → Privacy & Security and click Open Anyway, or run:
```
xattr -dr com.apple.quarantine /Applications/Yuzu.app
```
Let it fetch a model. Yuzu asks uzu which model fits your Mac and downloads it, with progress in the model picker. If it stops, click the model to try again.
Say hello. After that first download, Yuzu runs fully offline.

Requirements

Apple Silicon Mac (M1 or later)
macOS 26.4 or later
16 GB of RAM or more recommended
Several GB of free disk space for the model

Build from source

git clone https://github.com/adityachauhan0/yuzu
cd yuzu
open App/Yuzu.xcodeproj

The lid is open.
The model is yours.