eduwxyz/tucano

Your team of AI agents, on your Mac, Windows or Linux. Agents that talk in channels, write code, review each other and use your tools, on the Claude Code, Codex, OpenCode or Pi you already have. Anything that touches the real world waits for your yes.

Tucano is a free, open-source desktop app that lets you run a team of AI coding agents on your own computer, under your control. Instead of juggling separate terminals and copy-pasting between sessions, you get a chat-like workspace where agents talk in channels, assign work to each other with @mentions, and ask for your approval before doing anything that affects the real world.

What it does

  • Agents as a team – Each agent has a name, role, tools, and memory. They can investigate, fix, and review each other’s work, and you see every step in the conversation.
  • Your rules – You decide exactly what each agent can touch: no access, read-only, or read-and-write per tool. Anything like merging, purchasing, posting, or deleting waits for your explicit approval, which you can give from your computer or your phone (via Telegram).
  • Runs on what you already have – Agents think using AI harnesses you already use: Claude Code, Codex, Cursor CLI, Grok, OpenCode, or Pi. You keep your own subscriptions and logins; Tucano never sees your passwords or charges per token.
  • Everything stays local – Conversations, memory, logins, and files live only on your computer. There’s no Tucano server.

Key features

  • Channels – Create a channel, add agents, and ask. A router picks who answers, or you can @mention someone. Agents pass turns to each other, and every action is logged with token/time receipts.
  • Connectors – Connect any MCP-compatible tool (Jira, GitLab, GitHub, Slack, Linear, Figma, Notion, Datadog, Sentry, or your company’s internal server). Then set per-agent permissions.
  • Approvals – Write actions that matter show you the exact call and arguments; you can approve once, always allow a safe one, or refuse with a reason the agent reads.
  • Memory & skills – Agents remember your preferences and project details, and you can review, edit, or delete those memories. They can also propose reusable “skills” (written recipes) that the whole team can follow.
  • Routines – Schedule work on a calendar (e.g., “every Monday at 9am”) or trigger it when something new appears in a connector (new issue, failing monitor, email). Quiet hours keep the night quiet.
  • Tuca – A built-in helper that drafts new agents based on your description.
  • Usage & limits – See each agent’s monthly token usage against its limit, plus a weekly report of what each issue cost.
  • Computers & browsers – Give an agent an isolated Linux computer (via Docker) or your own computer, and a browser (Chrome) you can watch live and take over.
  • Your own apps – Agents can read and use apps on your computer (Notes, Calendar, Spotify) with read-only or read-and-use permissions, via the open-source agent-desktop project.

How it works

  • The app has a main window (React) and a core process that handles SQLite, runs, memory, and routines.
  • Each reply launches a harness CLI (like claude or codex) in an isolated environment, with no built-in tools and no saved session.
  • Tucano builds the context each turn: workspace, agent instructions, memories, skills, and recent conversation.
  • A local MCP gateway is the only source of tools for the CLI. Each turn gets a secret URL listing only that agent’s tools; any other call is refused. Sensitive calls become approval requests.

Installation

  • You need a Mac (Apple silicon or Intel), Linux (x64 or ARM64), or Windows 10/11, plus at least one harness installed and signed in (Claude Code is quickest).
  • Download the appropriate installer from the releases page (.dmg for Mac, .deb or AppImage for Linux, .exe or PowerShell command for Windows).
  • The first run walks you through setup: language, your name, which harnesses to use, models, tools, skills, and your first agent.
  • Tucano updates itself automatically from version 0.1.2 onward.

Your data

  • Everything is stored in a local folder (~/Library/Application Support/Tucano on Mac, ~/.config/Tucano on Linux, %APPDATA%\Tucano on Windows).
  • Connector passwords and tokens are encrypted with your system’s keychain/keyring before being saved.

Development

  • The project is open source (Apache 2.0). You can clone it, run npm install and npm run dev (requires Node 24+ and a signed-in harness). There’s also a fake-engine mode for testing without a subscription.

For more details, see the full README on GitHub.

Related

  • Project
  • Project
  • Project
  • Project