eduwxyz/tucano

What it is

Tucano is an Electron desktop app that organizes several AI coding agents into named teammates with roles, instructions, and per-tool permissions. Instead of juggling terminals and copy-pasting between sessions, you create channels, add agents, and mention them with @; a router can also pick who answers. Agents hand work to each other through mentions — for example, one investigates, another fixes, a third reviews — and each step is visible in the conversation with its tool calls and token/time receipts.

Agents do not ship their own model access. Each one wraps a harness already installed on your machine — Claude Code, Codex, OpenCode, or Pi — so it uses your existing subscription, company gateway, or configured providers. Each agent can choose a different harness and model, letting a team mix them.

Key capabilities

• MCP connectors: add remote (OAuth or token) or local MCP servers such as Jira, GitLab, GitHub, Slack, Linear, Figma, Notion, Datadog, Sentry, or an internal server, then grant access per agent and per tool as no access, read, or read/write.
• Approvals: write actions that matter are held with the exact call and arguments; you can approve, always allow, or refuse with a reason the agent reads. Approvals can also arrive on your phone via Telegram.
• Memory and skills: agents propose memories and reusable skills that you can review, edit, or forget.
• Routines: schedule work by date/time or connector events such as a new issue or a failing monitor.
• Isolated computers and browsers: give an agent a Docker-based Linux environment or your own machine, plus a dedicated Chrome you can watch and take over.
• Local-first data: conversations, memory, logins, and files stay on your computer; there is no Tucano server. Connector secrets are encrypted with a key held in the OS keychain/keyring.

Quick start

Prerequisites: a Mac with Apple silicon, an x64 or ARM64 Linux PC, or Windows 10/11 x64, plus at least one harness installed and signed in. On Windows you also need Git for Windows. Docker is optional for an agent's Linux computer, and Chrome/Chromium is optional for a dedicated browser.

The shortest path is Claude Code:

curl -fsSL https://claude.ai/install.sh | bash
claude auth login

Then install Tucano from the latest release: on macOS, open the .dmg and drag it into Applications; on Ubuntu/Debian, install the .deb with sudo apt install ./tucano-linux-amd64.deb; on other Linux distributions, use the AppImage (chmod +x first). On Windows, either run the PowerShell installer from the README or download tucano-win-x64.exe and pass SmartScreen via More info → Run anyway, since it is not code-signed. macOS and Linux also support a shell installer.

On first launch, pick English or Portuguese, then work through setup: your name, which harnesses agents may use (each has a Test now button), routing/memory models, what agents should know about you, optional connectors and skills, and your first agent. Setup ends in a #general channel where the agent is ready to answer.

Constraints and practical notes

• The app requires at least one supported harness; without it, agents cannot run. On Windows, Git Bash is required for agents' commands.
• macOS Apple silicon and Windows 10/11 x64 are the supported desktop targets; Linux is x64 or ARM64. The README tested macOS 26 and Ubuntu with KDE/Xfce.
• The Windows installer is unsigned, so SmartScreen warns on first run, and company-managed PCs may block unsigned apps. macOS builds are signed and notarized.
• On Ubuntu 24.04+, prefer the .deb because Electron's sandbox needs its AppArmor profile.
• On Linux, connector passwords are only stored in an open system keyring (GNOME Keyring or KWallet); without one, Tucano does not save them.
• Running on a headless server and using it from a browser is experimental.
• Licensed under Apache-2.0; the apps feature relies on Vercel Labs' separately installed agent-desktop.

HotRepo's take

The repository is new and small: the recorded daily window on 2026-11-01 shows 106 stars and 19 forks, with a single one-hour growth window gaining 5 stars and no fork change, and no GitHub daily rank has been recorded. That is not enough to support a trend claim in either direction.

What is notable from the material is the architecture rather than the launch numbers: agent turns run through the harness CLI with a per-turn secret MCP gateway URL that exposes only that agent's assigned tools, and sensitive calls become approvals before execution. That design is worth reading for anyone building a multi-agent desktop tool on top of existing coding harnesses; whether the permission boundary holds in practice would need independent review, which the recorded data does not provide.