DEV Community: Bellows AI

Qwen Code edits your files without asking. Here's the one-line fix.

Bellows AI — Thu, 05 Nov 2026 15:46:12 +0000

If you've installed Qwen Code recently, it's probably running in Auto mode.

That's the out-of-the-box default now. Unless you pass a flag or set it in your settings, every new session starts in Auto. In the source, it's the mode a normal session falls through to when nothing else is set.

Auto is a reasonable idea. It's also worth knowing exactly what it does before you point it at a repo you care about.

What Auto actually approves

Qwen Code's own docs describe three layers:

File edits inside your project are approved straight away. They don't even reach the safety check. The exceptions are Qwen's own config, .git/, package.json, CI workflows and a few similar files, which do get checked.
Read-only tools are approved straight away: reading files, search, listing, and read-only shell commands like ls, cat and git log.
Everything else goes to a classifier, which is a call to a language model (your configured fast model, or your main model if you haven't set one). That covers other shell commands, web fetches, sub-agents, MCP tools and edits outside the project. If the classifier says it's safe, it runs.

So in Auto, the agent can rewrite any source file in your project without a prompt, and whether a shell command runs is decided by a model. If your model is a small free one, that's the model making the call.

For a long unattended task in a throwaway branch, that may be exactly what you want. For anything else, you probably want it to ask.

The modes

Mode	Edits	Shell commands
`plan`	Not made: read-only analysis	Not run
`default` (shown as Ask Permissions)	You approve each one	You approve each one
`auto-edit`	Approved	You approve each one
`auto` (the default)	Approved inside the project	A classifier decides
`yolo`	Approved	Approved

The mode called default isn't the default any more. Its label in the app is now Ask Permissions, but the value you type is still default.

Make it ask: for this session

Press Shift+Tab (Tab on Windows) to cycle modes; the status bar shows where you are. Or type:

/approval-mode default

That lasts until you quit. The next session starts in Auto again.

Make it ask: every time

Put this in ~/.qwen/settings.json for every project, or .qwen/settings.json in one project:

{
  "tools": {
    "approvalMode": "default"
  }
}

Or pass it when you start:

qwen --approval-mode default

The flag wins over the setting, so you can keep default in your settings and still start a one-off session with --approval-mode auto-edit when you want speed.

If you run Qwen Code in an editor or app

When Qwen Code runs inside another app over the Agent Client Protocol (Zed, JetBrains, and other ACP clients), the same rule applies: each new session starts in Auto unless the app or your settings change it. The app can switch the mode, and many show a picker. If yours doesn't, put --approval-mode default in the agent's arguments, or use the settings file above. Both work over ACP; I checked.

I work on Bellows, a free desktop app that runs Qwen Code, Claude, Codex, Gemini CLI and other ACP agents. I found all this while adding Qwen Code to it. Bellows switches Qwen Code to Ask Permissions when it starts a session, shows each edit and command as a card to allow, and keeps every change as a diff you can put back.

Asked for on 2 October, shipped on 8 October: handing a coding-agent session to a teammate

Bellows AI — Sun, 11 Oct 2026 14:43:10 +0000

On 2 October someone opened an issue on Anthropic's tracker asking for a one-command way to hand a coding-agent session to a teammate (anthropics/claude-code#98937). Everyone working with agents knows it: you dig into something for an hour, the agent has all the context, and then someone else has to pick it up. Forwarding a chat log doesn't give their agent that context.

Six days later it shipped in Bellows 1.19.0. Here's what it does, and why we could move that fast.

Hand over: one file, any agent

Bellows is a free desktop app for Windows, macOS and Linux that runs coding agents: Claude, Codex, Gemini CLI, Qwen Code, any ACP agent, and local models.

In any session, Hand over saves one .bellows-session file:

the whole conversation, with every tool call and its result;
a handover the agent writes as you hand over: where things stand, what's next, what to watch out for;
keys and passwords redacted before the file is written;
a digest, so a file changed after it was made says so when it's opened.

Your teammate double-clicks it. Bellows shows who it's from and what it holds. They pick their own checkout of the repo and an agent, and choose Carry on. Their agent starts from your handover and the last exchanges, and the whole conversation is in their History.

And it doesn't have to be the same agent. You worked in Claude; they can carry on in Codex.

The rest of 1.19.0

Hand-over wasn't the only thing in the release:

Review edits live. A Review pane shows each edit as a highlighted diff the moment the agent asks for it: proposed, then applied or not.
Answer prompts from anywhere. Pair your phone once and every permission prompt from every session reaches it on any network, sealed end to end, to answer Allow once or Deny. The web console works too.
Switch agents mid-task (1.17.0). Continue with… hands the task to another agent; the one you leave writes a handover and the project remembers.

Pairing, live

The feature that started all this is live sharing. Send a colleague a link to the session you're running. They watch from a browser, nothing to install. Each link carries one role: watch, suggest or type. Suggestions wait in a queue and reach the agent only when you accept them. It was asked for publicly too: anthropics/claude-code#60082 and openai/codex#46016.

Why it shipped in six days

We ship small and often: eight releases between 5 and 8 October. Tell us what you need and it goes to the top of the list.

Try Bellows, free, and tell us what you'd want next in the comments or on our issue tracker. We act fast on feedback.

I use three coding agents. So I built one desktop app to run them all, under one set of rules

Bellows AI — Mon, 05 Oct 2026 17:43:02 +0000

Most days I use more than one coding agent: Codex for one job, Gemini CLI for another, Claude for a third. Each lives in its own terminal, asks permission its own way, and leaves no shared record of what it was allowed to do.

So I built Bellows: a free desktop app for Windows, Mac and Linux that runs them all side by side, under one set of rules.

Every agent, one window

Run Codex, Gemini CLI and Claude sessions in tabs, in the same project. Gemini CLI, Codex and Qwen Code have presets, and Bellows is an Agent Client Protocol (ACP) client, so any agent that speaks ACP is one command away. If Ollama or LM Studio is running, one click starts a session on a local model.

One set of rules, one record

Permission prompts are buttons, and they look the same whichever agent asks. Save an answer as a rule and it applies to every agent. Every allow and deny goes into a hash-linked audit trail, so an edit or a gap shows.

Codex runs through its own app-server, the one mode in which it asks before commands and edits. The few commands it treats as safe still run unasked, and Bellows records each one as having run.

Pair on a live session

Send a colleague a link to your running session. Their link carries one role: watch, suggest or type. Suggestions wait for you to accept them, and answering permission prompts is a separate switch, off by default. Everything stays on your machine. On your own network it's free.

Keep what worked

When an agent gets a file job right, press Keep as a tool. It becomes a small app with a Run button: next month, pick next month's files and press Run, with no agent and no AI charge. Kept tools run sealed off (no network, no shell), can be sent to a colleague as one file, and Check proves they still work.

Room to experiment

Start any agent in its own git worktree and review what it changed before you keep it. Resume or fork old sessions, and rewind files in Claude sessions.

What it doesn't do (yet)

No model included: each agent signs in its own way, and Claude runs on your Anthropic API key.
The Windows download isn't code-signed yet (SmartScreen: More info, then Run anyway). Mac needs Apple silicon.
Several agents on one job at once, with a side-by-side review, is what I'm building next.

The desktop app is free, all of it. A paid Team plan adds one policy for every machine and a shared audit trail.

Running agents side by side: https://bellowsai.app/guides/run-several-coding-agents-side-by-side

Which agents do you use together, and what would get you out of the terminal?