Sidekicks `1.3.0` is out 🎉

One week after the first blog post, here is already a new version of Sidekicks. On the menu for this 1.3.0: better file management, multiple terminals that keep their state, agents that talk to each other in a channel (I promised you 😉, but be careful, it's a 🚧 WIP and how well it works will depend on the "intelligence" of your models), the ability to unload a local model from memory, and above all Claude, Codex, Gemini, Docker Agent and mini-me running in an sbx sandbox in a few clicks.

Managing an agent's files, like in a file manager

Until now, the Files tab was mostly used to browse an agent's files. Now, you can really work with them:

• select one or more files (⌘/Ctrl-click, ⇧-click),
• rename them with ✏️ or a double-click,
• move them with drag and drop, or with Move to…,
• delete them all at once, or create a folder.

And of course, a right-click offers the same actions. Don't worry about protected files: they stay protected 🔒.

Another small improvement: source code is now syntax highlighted, for about thirty languages (Go, Python, JavaScript, TypeScript, Rust, Java, shell, YAML, SQL…), and the editors follow the theme chosen in Sidekicks. Finally, the 📂 icon, next to each file and above the open file, shows it directly in the Finder (or in the Explorer on Windows - I haven't tested Sidekicks on Windows yet, I'm trying to do that soon).

Terminals that keep running (and several per agent)

It was a bit frustrating: you ran a command in an agent's terminal, you went to take a look at a channel.. and when you came back, everything was gone 😩. That's over: a terminal keeps its content and whatever is running in it when you open another agent, a channel or the settings.

And while we're at it, the ＋ button opens another terminal for the same agent, each one in its own tab. × closes one. The Terminal tab and the Agents panel show you how many are open.

Agents calling each other in a channel

This is my favorite new feature 🤓 (actually it comes from a need I had for my work). In a channel, when an agent's answer mentions another member with @name, that member answers right after, knowing who is talking to it. You just have to ask:

@alice plan a trip to Lyon, and ask @bob for the trains

Even better, a message addressed to several agents gives each one its share of the work. With:

@alice write a story, then @bob summarize it

Alice receives "write a story", and Bob receives "summarize it"… after Alice's answer, of course.

✋ Important: to keep your agents from spending the whole day passing the ball back and forth, a message goes from agent to agent 3 times at most. You can change this with the limits.maxAgentHandoffs key (0 disables the feature), and the ■ button stops the chain at any time.

limits:
  maxRunningAgents: 2
  stopIdleAgentsAfter: 10m
  maxAgentHandoffs: 3

Unloading a model from memory

When you run several small models locally, memory pressure goes up quickly and it will slow your machine down dramatically, or even freeze it 🙀. In the Log panel, next to the memory pressure, you will now find ⏏ Unload the model (for the displayed agent) and ⏏ Unload all. The ⏏ button is also available on an agent's row and in its header.

The agent keeps running and its conversations don't change: the model will simply be reloaded on the next message.

This only works with agents that offer the /unload command. That's the case of mini-me (mm) 😉

Claude, Codex, Gemini, Docker Agent and mini-me in an sbx sandbox

In the previous post, I told you we would soon see how to run our agents in a totally secure way with Docker sbx. Well, it has become much simpler 🎉.

In the add-agent wizard, choose In an sbx sandbox, then drop the sandbox's sbxenv.yaml file. The sandbox, the folder and the commands are filled in automatically, and it's the file's agent: line that chooses the program to launch:

• claude-agent-acp for Claude,
• codex-acp for Codex,
• gemini --acp for Gemini,
• docker-agent serve acp .../agent.yaml for Docker Agent,
• mm --acp --config .../agent.yaml for mini-me.

(For Docker Agent and mini-me, agent.yaml is the agent file in the sandbox's workspace.)

The wizard also protects you from classic mistakes: it refuses an sbxenv.yaml made for another agent than the chosen program, as well as a Docker Agent without its agent file.

To get started quickly, you'll find ready-to-use examples in agents/snippets: claude-acp, codex-acp, gemini-acp, docker-agent-acp and mini-me-acp.

✋ Important: your API keys are "managed" by sbx (sbx secret set ..), never stored in a file and never read by the agent.

Good to know before updating

• On macOS, drag the new Sidekicks into Applications and confirm the replacement: your configuration and your history are kept. The macOS packages are signed and notarized.
• Your configuration files work as they are. The only new key, limits.maxAgentHandoffs, is optional (3 by default).
• The Claude, Codex and Gemini kits are provided by third parties, not by Sidekicks. Be careful with Codex: you now need the v3 kit docker.io/sbx/codex-acp, the old docker.io/sbx/codex-acp-kit no longer works with the sbx codex agent.

Where to download?

The packages are available here.

That's it for this 1.3.0, see you very soon 👋