Tucano

Your team of AI agents, on your Mac. Your rules.

Agents that talk in channels, write code, review each other and use your team's tools.

They do the work. Anything that touches the real world waits for your yes.

English · Português · tucano.sh · @tucanosh

https://github.com/user-attachments/assets/533e1bf1-b4ca-4f4a-aad4-e015c6b4c421

Why Tucano

Coding agents got good. Working with several of them is still a mess of terminals, copy-paste between sessions, and hoping nothing runs that shouldn't. Tucano puts them in one place that feels like your team's chat, and gives you the controls a team lead would want.

• A team, not a prompt box. Agents with names, roles, tools and memory. They talk in channels, call each other with @mentions and hand work over: one finds the bug, another fixes it, a third reviews it, and you get the result.
• Your rules. Each agent sees only the tools you give it, with no access, read, or read and write per tool. Merges, purchases, posts, emails and deletes become an approval tied to the exact action, on your Mac or your phone.
• Runs on what you already have. Agents think with the Claude Code, Codex, OpenCode or Pi on your Mac, the way you set them up: your subscription, your company's AI gateway, or your providers. No API key to paste, no per-token bill from us, and Tucano never sees your password.
• Everything stays on your Mac. Conversations, memory, logins, files. There is no Tucano server.

What it does

Channels where agents work together

Create a channel, add the agents, and ask. Mention someone with @, or don't, and a router picks who answers. Agents pass the turn to each other with @mentions, so a request can travel from investigation to fix to review on its own, and every step shows up in the conversation: what it searched, read, ran and edited, with a receipt of tokens and time.

Connect any tool, even your company's internal ones

Tucano speaks MCP. Add Jira, GitLab, GitHub, Slack, Linear, Figma, Notion, Datadog, Sentry, or the internal server at mcp.yourcompany.internal: remote servers with OAuth or a token, or local ones. Then decide, per agent and per tool, what it can touch.

You decide how far each one goes

Write actions that matter wait for your OK, with the exact call and its arguments in front of you. Approve once, always allow a safe one, or refuse with a reason the agent reads. Approvals also reach your phone through Telegram. Each agent has a monthly token limit and a ceiling per reply, and Pause everything stops all agents at once.

It learns as you use it

Agents remember what matters about you, your projects and your preferences, and you see, approve, edit and forget every memory. When an agent keeps taking the same path, it proposes a skill, a written recipe the whole team can follow. Skills you already use in Claude Code and Codex come in with one click.

Routines, on a calendar

Schedule work for a day and time, or for when something new shows up in a connector: a new issue, a failing monitor, an email. Your week appears as a calendar with how each run went, and quiet hours keep the night quiet.

Tuca builds agents with you

Missing someone on the team? Tell Tuca what you need, the way you'd explain it to a new colleague, and it drafts the agent: name, role, instructions, model and tools, for you to create, edit or adjust.

Every answer with a receipt

Usage and limits shows each agent's month against its limit and your subscription's windows. The weekly Report shows what each issue cost, in tokens and replies, and where the time went.

Computers and browsers of their own

Give an agent a computer, an isolated Linux in Docker where it installs, runs and tests, or, if you choose so, your Mac. Lend it a folder or a repository. Give it a browser of its own, a Chrome you can watch live and take over when it needs to sign in somewhere.

Runs on what you already have

Each agent picks its harness and model, so a team can mix them: Opus to fix, GPT to review, Haiku to route.

Behind a company gateway? If your Claude Code or Codex already sends requests to your company's AI proxy, Tucano's agents go the same way, with no extra setup. Telemetry and error reports of a harness on a gateway stay off, and the model list comes from the gateway itself. To use only the gateway, uncheck the other harnesses in Settings → Models.

Install

You need a Mac with Apple silicon (M1 or later; tested on macOS 26) and at least one harness, installed and signed in. Optional: Docker Desktop to give agents a Linux computer, and Google Chrome for their browser (without it, the built-in browser is used).

1. Install a harness. Claude Code is the quickest:

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

Prefer Codex? npm install -g @openai/codex. You'll sign in from inside Tucano.

2. Install Tucano. Download the .dmg, open it and drag Tucano into Applications. The app is signed and notarized by Apple, so it opens with no warning.

Or from the Terminal, which also updates it later:

curl -fsSL https://github.com/eduwxyz/tucano/releases/latest/download/install.sh | sh

It downloads the latest release, puts Tucano in Applications and opens it. Run the same command to update.

3. The first run is the setup. Tucano speaks English and Portuguese; pick one on the first screen, or later in Settings. Then, in a few steps: your name; which harnesses agents may use (each with a Test now button that sends a real question); the models that route channels and keep memory; what every agent should know about you; your tools (optional); your skills (optional); and your first agent, from a template or from scratch. It ends in a #general channel with your agent already working.

The first time Tucano saves a connector password, macOS asks whether it may use the Keychain: click Always Allow.

Using it

• Connect a tool: Connectors → Add, paste the MCP server's address. OAuth servers open the browser; the others ask for a token. Then, on each agent's page, set its access to each connector.
• Archived agents leave the sidebar and stop answering, but keep their history and receipts. Find them under Archived to bring them back.
• Approve from your phone: connect the Telegram bot in Settings → Phone and GitHub; approvals and questions arrive with their buttons.
• Getting started, in the sidebar, walks you through the rest: build a team, connect a tool, teach the memory, schedule a routine, lend a project.

How it works

flowchart LR
  subgraph Mac["Your Mac"]
    UI["Window<br/>(React)"] <-->|MessagePort| Core["Core process<br/>SQLite · runs · memory · routines"]
    Core -->|"one turn, isolated"| CLI["Harness CLI<br/>claude · codex · opencode · pi"]
    CLI -->|"secret per-turn URL<br/>only this agent's tools"| GW["Tucano MCP gateway"]
    GW --> Tools["Connectors (MCP)<br/>computer · browser · memory"]
    GW -.->|"write action"| OK{{"Your approval"}}
  end
  CLI --> Model["Your subscription<br/>or company gateway"]

• Main process (Electron): the window, notifications, the agents' browsers, the Keychain.
• Core (a utility process): SQLite, harnesses, runs, connectors, memory and routines. No agent work runs in the window.
• Each reply opens the harness CLI isolated from your personal configuration, with no built-in tools and no saved session. Tucano builds the context every turn: the workspace, the agent's instructions, its memories and skills, and the recent conversation.
• The gateway is a local MCP server and the CLI's only source of tools. Each turn gets a secret URL listing just that agent's tools; any other call is refused. Sensitive calls stop there and become an approval, which runs exactly the call you saw.

More on the permission model in SECURITY.md.

Your data

Everything lives in ~/Library/Application Support/Tucano/: the database (tucano.db), a daily backup in backups/, each agent's computer folder and the files agents show you. Connector passwords and tokens are encrypted with a key in the macOS Keychain before they reach the database.

1. Quit Tucano (⌘Q) and delete /Applications/Tucano.app.
2. To delete your data too: rm -rf ~/Library/Application\ Support/Tucano.
3. If you used the Linux computer, delete the tucano-* containers in Docker Desktop.
4. Optionally, delete the Tucano Safe Storage item in Keychain Access.

Claude Code shows as "not installed".
Tucano looks for claude in the same PATH as your Terminal. Run which claude: if nothing comes back, the install didn't finish. If it does, quit Tucano (⌘Q) and open it again.

Claude Code shows as "not signed in".
Run claude auth login in the Terminal, then click Check again in Settings.

An agent stopped at its limit.
Each agent has a monthly token limit. Raise it on the agent's page, or see the usage in Usage and limits.

"Docker isn't running" when starting the Linux computer.
Install and open Docker Desktop. Without it, choose None or This Mac as the agent's computer.

A connector needs authorization.
Click Authorize in the browser, on the agent's page or in Connectors, and sign in to the tool.

macOS asks for the Keychain after updating from 0.1.0.
0.1.0 wasn't signed with Tucano's Apple certificate; later versions are, so it asks only this once. Click Always Allow.

Development

git clone https://github.com/eduwxyz/tucano.git
cd tucano
npm install
npm run dev

You need Node 24+ and one harness signed in. npm run dev keeps its own data in ~/Library/Application Support/Tucano Dev, apart from an installed Tucano. With TUCANO_FAKE_ENGINES=1, agents answer with a scripted engine and a sample team, with no subscription at all.

Read CONTRIBUTING.md before opening a pull request. The original design document, in Portuguese, is in docs/spec-mvp.md.

License

Apache-2.0. Tucano is not affiliated with Anthropic, OpenAI or any of the tools it connects to.