MAO vs Nimbalyst, formerly Crystal: terminal agents on a canvas or agent sessions beside visual editors?

Nimbalyst, the app that replaced Crystal, is a free, MIT-licensed desktop workspace for macOS, Windows and Linux where you run Claude Code and Codex sessions in a chat-style agent window, each optionally in its own git worktree, next to visual editors for Markdown, mockups, diagrams and code, a task tracker and an iOS companion app. MAO is a paid desktop app that runs each coding agent's own CLI as a real terminal on a canvas, connects agents so they can send each other tasks, and adds browser and device portals and remote workspaces over SSH. Choose Nimbalyst if your work centres on documents, plans and reviewing agent edits visually; choose MAO if it centres on many terminal agents working and handing off together.

What is Nimbalyst?#

An open-source visual workspace for coding agents, and the successor to Crystal.

Crystal, by stravu, was a desktop app for running several Codex and Claude Code sessions in parallel git worktrees. Its repository now says Crystal was deprecated in February 2026 and replaced by Nimbalyst; you can keep using Crystal, but its makers recommend Nimbalyst for active updates.

Nimbalyst's README calls it an open-source visual workspace for building with Codex, Claude Code, OpenCode and other coding agents. Its main parts:

• Agent sessions. A chat-style agent window shows each session's transcript, tool calls and streamed output. You run sessions in parallel, search and resume them from the sidebar, group them into workstreams and track them on a kanban board.
• Worktrees. With Developer Mode on, New Worktree starts a session on a new branch in its own folder; you review, commit and merge the branch from the panel beside the session.
• Visual editors. WYSIWYG editors for Markdown, mockups, Mermaid, Excalidraw, CSV, data models and code (Monaco). The agent's edits appear as red and green diffs that you accept or reject one by one.
• Task tracking. Plans, bugs, features and todos that agents read and update with you.
• Agent Teams and Super Loops. A session can spawn teammate agents for independent subtasks and waits for them to finish; a Super Loop runs one agent repeatedly on a task in its own worktree.
• Browser. An in-app browser for HTML files and live URLs that the agent can inspect, click, type into, scroll, screenshot and run JavaScript in.
• More. A Ghostty-powered terminal, Git tools with AI-drafted commits, an MCP client, an extension SDK, and an iOS app that pairs with the desktop to reply to agents, review diffs and get push notifications.

Nimbalyst is built with Electron. The app is free for individuals with no account; a paid Wiki Teams plan adds a shared team wiki.

How do MAO and Nimbalyst differ?#

Nimbalyst wraps agent sessions in editors, trackers and a chat window; MAO keeps each agent's own terminal and connects agents to each other on a canvas.

When does Nimbalyst fit better?#

When the agents' work is documents, plans and designs as much as code, or when you want it free and open source.

• You want open source at no cost. Nimbalyst is MIT-licensed and free for individuals, with no feature limits and no account. MAO is closed source and a one-time purchase.
• You review agent edits visually. Agents edit Markdown, mockups, diagrams, CSV and data models in WYSIWYG editors, and you accept or reject each change as a red and green diff.
• You plan and track work next to the agents. The task tracker and the session kanban link plans, bugs and sessions to the files they touched.
• You answer agents from your phone. The iOS app shows which sessions need you, takes text or voice replies, and sends a push notification when an agent is waiting.
• You want Cursor Agent or GitHub Copilot. Both are Nimbalyst providers in alpha; neither is a listed MAO provider.
• You want one agent to fan out by itself. Agent Teams lets a session spawn teammates for subtasks, and Super Loops run one agent repeatedly on a task, with fresh context each time.

When does MAO fit better?#

When you want each agent's real terminal, agents that hand work to each other, devices to test on, or a remote machine.

• The agent runs as its own CLI. Each node is the provider's CLI in a real terminal, with its own login, model and configuration, so it behaves as it does in your terminal. A chat view is there when you want it, and hands the session back to the terminal.
• Agents hand work to each other. Draw a connection and one agent can send a task to another, for example from Claude Code to Codex, and read the result by task ID.
• Agents test on devices. Besides a browser portal, device portals put an Android emulator or phone, or the iOS Simulator on macOS, on the canvas for a connected agent to drive.
• The project lives on a server. A remote workspace keeps the repository, terminals and agents on a Linux machine over SSH, or inside WSL, while the canvas stays on your desktop.
• Every agent is visible at once. Agents, notes, files and portals sit side by side on one canvas, and zooming out turns them into status cards.
• Native app, no telemetry by default. MAO is a desktop-native Rust app with a GPU-rendered GPUI canvas, and sends no telemetry unless you opt in. Nimbalyst's README says it sends anonymous usage analytics, which you can turn off in its settings.

How do I compare them on a real task?#

Give both the same task with a planning step and two independent parts, and see where the work stays easy to follow.

1. 1Pick a change with a short plan and two independent parts, such as an API endpoint and its tests, in a repository you can throw away.
2. 2In Nimbalyst, write the plan in a Markdown document, start two sessions with New Worktree, and review each session's diff before merging it from the session panel.
3. 3In MAO (the 7-day trial is enough), put the plan in a note, create a layer, place a Claude Code and a Codex agent on it, connect them, and have one ask the other to review its work with mao ask.
4. 4Land the layer into Ground and compare: how often you relayed messages between agents, where you reviewed the diff, and how easy it was to see which agent needed you.

Frequently asked questions#

Sources#

• Crystal README: Crystal is deprecated (February 2026) and replaced by Nimbalyst; you can still use it; the repository description of Crystal as parallel Codex and Claude Code sessions in git worktrees.
• Nimbalyst README: features, supported agents, platforms and requirements, the iOS companion, telemetry and its opt-out, Electron, and the MIT licence.
• Nimbalyst licence and LICENSING.md: MIT, and the sync server as a separate project.
• Nimbalyst website: the app and the wiki, downloads for macOS, Windows and Linux, free for individuals.
• Nimbalyst pricing: the free app, no account, bring your own AI access, Wiki Teams price.
• Agent window and session management: the chat-style transcript, parallel sessions, workstreams and the kanban.
• Worktrees: New Worktree, Developer Mode, merging from the session panel.
• Agent Teams and Super Loops: teammate agents and repeated runs in a worktree.
• Browser: the in-app browser and what the agent can do in it.
• AI provider setup: the provider list, which are alpha, and how Codex signs in.
• Terminal window: the Ghostty-powered embedded terminal.
• Mobile app: iOS, pairing with the desktop, work running on the desktop.
• MAO product facts: platforms, supported agents, price, free trial, telemetry, native Rust and GPUI.
• Run a shell or agent and Use a structured ACP chat: terminal nodes, provider logins, the optional chat view.
• Create a worktree layer and Land a layer into Ground: layers as Git worktrees and the checks before landing.
• Send a task and collect its result: mao ask, mao send, mao next and mao result between connected agents.
• Browser and device portals and View an iOS Simulator portal: browser, Android and iOS Simulator portals.
• Remote workspaces: Linux servers over SSH, WSL, and that remote workspaces come with the licence.

Checked on 2026-10-28.