AionUi

You will like this if you run three CLI agents in three terminals and lose track of which one wrote the file. The daily trait is detection: it finds the agents already on your machine, Claude Code, Codex, Qwen Code, Gemini CLI, Goose and a dozen others, and gives each a session with its own context beside a preview pane that shows the PDF, sheet or diff the agent just produced.

Where it hurts is the Electron desktop app itself, another thing to update and another place your keys live. Pick it for office work driven by coding agents. Pick Claude Squad if you would rather stay in tmux.

The risk is the shared folder. In Team Mode all agents read and write the same workspace while a leader delegates through a Team MCP Server, and the README notes that a teammate that goes quiet is escalated to failed and removed with one click. Two agents editing one file with no worktree between them is the classic corrupted-edit scenario, and the document describes no merge step.

The consequence: keep teams to read-heavy tasks, or give each teammate a copy. What it does right: every agent has its own permission dialog, and the sidebar shows a badge for pending approvals, so nothing lands without a named agent asking first.

The multi-agent design is described concretely. 1. A leader agent receives the instruction and splits it into subtasks. 2. Teammates run in parallel, each on its own model, connected through the Agent Client Protocol. 3. Coordination is an asynchronous mailbox and a shared task board rather than direct calls. 4. The built-in agent is an embedded engine written in Rust.

Verification is absent from the description; a subtask is done when the teammate says so. No benchmark is published for any mode. The observation: an unattended mode whose available modes and permission behaviour depend on the selected agent is a dozen modes, and the README says as much.

The README opens with a partnership: Kimi gives contributors a plan the campaign values at $39 a month, marketing spent in kind. The positioning is a comparison table against Claude Cowork at $100 a month and macOS only, and distribution is YouTube reviews with view counts in the README and articles on WeChat. There is no price and no hint of one.

Moat: none; the app hosts other vendors' agents. Pricing power: negative, the pitch is free. Likely pivot: office document generation for the Chinese market, where PPT and Word assistants are already the headline. Likely acquirer: a Chinese office-suite or model vendor. Position: short the company, long the idea.

The demo is a PowerPoint generated from a chat. For sixty seats it is a desktop install on sixty laptops, 4 GB of memory and half a gigabyte of disk each, plus whatever keys each engineer pastes in. Scheduled tasks run around the clock and the app prevents the machine from sleeping while they do, which is sixty laptops that never sleep. Remote access is a WebUI with QR or password login; there is no SSO, no SCIM, no audit trail, and every conversation sits in a SQLite file on the device.

Support is a Discord and a WeChat group. Onboarding is easy. Not yet.

Apache-2.0, but split across two repositories: the Electron front end here and the Rust engine in AionCore, where the built-in assistants live as a JSON catalog I can edit. Local models go in through a Custom platform entry pointing at Ollama or LM Studio, or through a NewAPI gateway if I want one door for everything. MCP management is centralised, and the app injects or syncs the transport each hosted agent can handle rather than making me configure them one by one.

Skills come in three tiers with an Extension SDK for the third, and the UI takes my own CSS. Forkable, if I fork both halves. Acceptable.