---
type: "post"
id: "tucano-0-2-2-lets-its-multi-agent-d6011d"
slug: "tucano-0-2-2-lets-its-multi-agent-d6011d"
title: "Tucano 0.2.2 lets its multi-agent team run headless on a self-hosted server and adds Gemini CLI as a harness"
description: "Tucano 0.2.2 adds an experimental server mode for running its multi-agent workspace on a machine the user controls, such as a VPS, while accessing the interface in a browser."
retrieval_nugget: "Tucano 0.2.2 adds an experimental server mode for running its multi-agent workspace on a machine the user controls, such as a VPS, while accessing the interface in a browser. It also adds Gemini CLI as a harness. For builders, this is a practical shift from a laptop-bound workspace toward an always-on, self-hosted agent runtime."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: ["gemini","claude","codex"]
entities: ["Tucano","Gemini CLI","Claude Code","Codex","OpenCode","Pi","Google"]
related_patterns: ["harness-architecture-outlives-model-choice","agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "Tucano 0.2.2 adds experimental self-hosted server mode and Gemini CLI support, distinct from prior assistant integrations and self-hosting releases."
prior_refs: ["memory:2d205efe-ff7b-87a9-ac7a-c6a7e11fbfc9","memory:ce875981-37bc-8218-a6d0-aa68ff2c05f8","memory:0acbb120-23c6-8ba3-ad08-e82af4d21370","memory:79ab635c-75d7-80e2-aa83-87862a2233ed"]
source_url: "https://github.com/eduwxyz/tucano/releases/tag/v0.2.2"
source_urls: ["https://github.com/eduwxyz/tucano/releases/tag/v0.2.2"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:tucano-0-2-2-lets-its-multi-agent-d6011d"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/tucano-0-2-2-lets-its-multi-agent-d6011d/","markdown":"https://newruntime.com/posts/tucano-0-2-2-lets-its-multi-agent-d6011d.md","json":"https://newruntime.com/posts/tucano-0-2-2-lets-its-multi-agent-d6011d.json"}
---

# Tucano 0.2.2 lets its multi-agent team run headless on a self-hosted server and adds Gemini CLI as a harness

## Retrieval answer

Tucano 0.2.2 adds an experimental server mode for running its multi-agent workspace on a machine the user controls, such as a VPS, while accessing the interface in a browser. It also adds Gemini CLI as a harness. For builders, this is a practical shift from a laptop-bound workspace toward an always-on, self-hosted agent runtime.

## Summary

Tucano 0.2.2 adds an experimental server mode for running its multi-agent workspace on a machine the user controls, such as a VPS, while accessing the interface in a browser. It also adds Gemini CLI as a harness. For builders, this is a practical shift from a laptop-bound workspace toward an always-on, self-hosted agent runtime.

## What happened

Tucano describes itself as a team of AI agents living on a user’s computer. Its agents work in chat-like channels, where they can research, write, code, and use company tools while the user chooses how much authority they receive. Rather than providing a proprietary model endpoint, Tucano works with coding-agent command-line tools that the user already has: Claude Code, Codex, OpenCode, Pi, and now Gemini CLI. Tucano released version 0.2.2 on October 31.

The main addition is an experimental server mode. Running `npm run server` starts Tucano without a desktop window on a machine owned by the operator, including a VPS, and serves Tucano’s own screens to a browser on another computer or phone. The release says routines can keep running while the user’s computer sleeps. The interaction surface remains Tucano’s interface, but the workspace no longer has to be attached to the active desktop.

The release specifies several controls and operating details. Every request requires the server token, and the server listens only on the server itself unless the operator chooses another configuration. Documentation covers access through an SSH tunnel, Tailscale, or a reverse proxy, plus a systemd unit. Secrets are sealed with a key that only the server user can read. OAuth connectors work in server mode by having the operator paste the final address from the provider sign-in flow back into Tucano. Agent-browser takeover uses a live picture in the panel; screenshots come from Chrome running on the server. On a private network such as Tailscale, pages served by agents can open in Tucano’s preview panel.

Version 0.2.2 also adds Gemini CLI as a harness, letting agents use Google’s Gemini models through the CLI to which the user is signed in. This extends, rather than replaces, the existing harness options. A new Install button opens the relevant installer in PowerShell on Windows, Terminal on macOS, or the desktop terminal on Linux; Tucano detects the harness after installation and points users to Node.js when required. The release also improves Windows discovery of Claude Code by checking installer locations such as `~/.local/bin` and npm’s global folder. Server mode remains explicitly experimental.

## What the archive adds

This is a new fact on a known story: an orchestration interface can sit above several coding-agent CLIs rather than being tied to one model vendor. [DeepSeek Harness ships dsh-v0.2.2-alpha.2 with plugin updates, six new UI languages and an experimental Codex/Claude CLI subagent tool](https://newruntime.com/posts/deepseek-harness-ships-dsh-v0-2-2-alpha-ca4e22/) similarly treated coding CLIs as components within a higher-level agent environment. Tucano adds the distinct step of operating that environment on a self-hosted server.

## Trend signal

This supports [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). Gemini CLI joins Claude Code, Codex, OpenCode, and Pi behind one team interface, while server mode lets that interface persist beyond a developer’s active desktop session. The release does not show that harnesses are interchangeable in quality, cost, or permissions, but it makes the orchestration layer more durable than any one model connection.

## Connections to other events

The earlier DeepSeek Harness release added an experimental subagent tool for Codex and Claude CLI. Tucano reaches a related boundary from another direction: it lets a multi-agent workspace choose Gemini CLI as a harness and operate away from the local desktop. In both cases, the command-line agent becomes a component rather than the complete product surface.

## Why it matters

Teams testing coding agents often choose between manual local sessions and a managed cloud product. Tucano introduces a third shape: run the orchestrator on infrastructure the team operates, retain existing authenticated coding CLIs, and reach the control interface through a browser. That can suit recurring research, coding, or connector-driven routines that should not stop when a laptop closes.

The trade-off is operational responsibility. Self-hosting changes where credentials, browser activity, agent screenshots, and network exposure live. Local-only listening, a per-request token, a server-user key for secrets, and private-network access are useful controls, but they do not decide which tools an agent may use or which actions need review. Treat server mode as a runtime deployment choice, not a convenience toggle.

## Limits and uncertainty

The source is Tucano’s release note and calls server mode experimental. It provides no benchmarks, uptime figures, concurrency limits, threat model, migration guidance, or evidence of team-scale operation. It does not establish feature parity among Claude Code, Codex, OpenCode, Pi, and Gemini CLI, or explain the pricing, quotas, and authentication constraints inherited from each CLI and subscription. The release describes sealed secrets and local-only listening by default, but does not independently verify the implementation or the safety of an operator’s tunnel, Tailscale configuration, or reverse proxy.

## Skill to build now

Take one recurring, low-risk agent routine and document its runtime boundary: the machine that executes it, account and CLI it uses, tools it can invoke, network route to its control panel, and person able to revoke access. If testing Tucano server mode, begin with local-only listening and a private access path such as SSH tunnelling or Tailscale. Treat the server token and secret-sealing key as deployment credentials.

## Reader takeaway

Tucano 0.2.2 is not a new model; it is infrastructure for a multi-agent workspace. It can now keep running on a user-controlled server and use Gemini CLI alongside other coding-agent harnesses. The opportunity is persistent agent work, while the requirement is deliberate operation of its runtime, access path, and permissions.

## Open question

How reliably do Tucano’s different CLI harnesses preserve permissions, tool behavior, and long-running routine execution in experimental server mode?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Codex](https://newruntime.com/topics/codex.md): Explore the codex topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)