TLDR
Install the ClawTab desktop app. It listens for Claude Code instances in tmux and sends keystrokes to automatically accept permission prompts. Enable auto-yes per-pane from your phone or the desktop app, disable it any time.
The Problem: Permission Prompts Stall Your Agents
Claude Code asks for permission before taking certain actions - writing files, running commands, installing packages. These are sensible guardrails. But when you're running a large refactoring job or a batch of agents across multiple panes, each permission prompt stalls the agent until you manually approve it.
If you're at your desk, it's a minor interruption. If you stepped away, it means you come back to an agent that's been sitting idle for 20 minutes waiting for you to press "1" for Yes.
Claude Code has a built-in --dangerously-skip-permissions flag that disables all permission checks entirely. There's also the newer auto mode with an AI safety classifier. But both are session-scoped and can't be toggled remotely. Other tools like VS Code with GitHub Copilot or Cursor handle this differently - they either auto-accept everything in agent mode or require you to sit at your desk clicking approve. Neither gives you selective remote control.
Auto-yes mode takes a different approach. It watches for permission prompts and automatically accepts them per-pane, so your agents keep making progress even when you're not watching - while other panes still require manual approval.
How It Works
When you enable auto-yes for a Claude Code pane, ClawTab's desktop app starts polling the terminal output at double the normal rate (every 500ms instead of every 2 seconds). When it detects a numbered permission prompt, it does two things:
- Looks for the best "yes" option - it prefers "Yes, during this session" over a plain "Yes", keeping the scope of the permission as narrow as possible
- Sends the option number as a keystroke directly to the tmux pane, just as if you typed it yourself
The detection uses the same question-parsing logic that powers the remote notification cards. It reads the terminal buffer, identifies numbered options, matches against known "yes" patterns, and sends the keystroke. No special API or Claude Code plugin required - it works by reading and typing into the terminal like a human would.
This is fundamentally different from how other AI coding tools handle permissions. JetBrains Junie and Windsurf run agents inside their IDE sandbox, which limits what the agent can do. Claude Code runs directly in your terminal with full system access, which means permission prompts actually matter. ClawTab's auto-yes gives you a way to selectively bypass them without disabling the safety system entirely.
Each answered question is tracked with a stable hash so the same prompt is never answered twice. Once a question disappears from the terminal output (meaning Claude accepted the answer and moved on), ClawTab clears its tracking state and is ready for the next one.
Enabling Auto-Yes From Your Phone
The most common way to enable auto-yes is from the mobile app. When a Claude Code agent asks a question and you get the notification card, you'll see option buttons at the bottom of the screen - the detected choices from the terminal prompt, plus a red "Yes all" button on the right.
Tapping "Yes all" does two things immediately:
- Answers the current question with "Yes"
- Enables auto-yes for that pane going forward
A confirmation dialog explains what's about to happen: all future permission prompts for that pane will be automatically accepted until you disable it.
Once enabled, you'll see a banner at the top of the jobs list showing which panes have auto-yes active, along with a "Disable" button for each one. The banner persists across app restarts and reconnections - the relay server caches the auto-yes state and replays it when your phone reconnects.
This remote-first workflow is what sets ClawTab apart from tools like Claude Code's native terminal experience or IDE-based agents like GitHub Copilot in VS Code. You don't need to be at your desk or have your IDE open. You can approve everything from your phone while walking the dog.
The Auto-Yes Banner
When auto-yes is active, a prominent banner appears at the top of the jobs list - both on mobile and on the desktop app. The banner shows the pane path (like "~/workspace") and a "Disable" button.
This serves as a persistent reminder that auto-approval is active. You won't forget it's running because it's always visible at the top of the screen.
Tapping "Disable" turns auto-yes off immediately. The next permission prompt will wait for your manual approval as usual.
Cross-Device Sync
Auto-yes state syncs between your desktop and all connected mobile devices through the relay server. Enable it from your phone, and the desktop starts auto-accepting within the next polling cycle. Disable it from the desktop UI, and the banner disappears on your phone.
The relay caches the current auto-yes state per user. When a mobile device reconnects (after a network drop, app restart, or switching devices), the relay replays the cached state so the UI stays accurate without needing to wait for the desktop to re-broadcast.
This also works with workspace sharing. If a colleague has shared access to your workspace, they can see which panes have auto-yes enabled and can disable it if needed.
Push Notification Suppression
When auto-yes is active for a pane, ClawTab suppresses push notifications for questions from that pane. There's no point buzzing your phone about a prompt that's going to be auto-accepted in half a second.
You still get push notifications for other events - job completion, failures, and questions from panes that don't have auto-yes enabled. Only the permission-prompt notifications are filtered out.
How ClawTab Compares to Other Approaches
There are several ways developers deal with AI agent permission prompts today:
- Claude Code
--dangerously-skip-permissions- disables all permission checks globally. Fast but removes all guardrails. You can't selectively approve one task while blocking another. - Claude Code auto mode (new, March 2026) - uses a safety classifier to auto-approve actions that match your task scope. Requires Team plan, Sonnet 4.6 or Opus 4.6, and is session-scoped. No remote control or per-pane granularity.
- Cursor agent mode - auto-accepts most actions within the IDE, but you must be at your desk to monitor. No remote approval, no per-task granularity.
- GitHub Copilot in VS Code - requires clicking "Accept" in the IDE for each tool call. No remote access, no batch approval.
- Cline / Aide - VS Code extensions that also require in-IDE approval. Some support "auto-approve" settings in config, but it's all-or-nothing per session.
- Devin - runs in a sandboxed cloud environment, sidestepping the permission problem entirely. But you give up local development and direct terminal access.
ClawTab's auto-yes is unique because it's per-pane, remotely controllable, and works with the existing Claude Code permission system rather than replacing it. You keep the safety prompts where you want them and auto-accept where you don't.
ClawTab Auto-Yes vs Claude Code Auto Mode
Anthropic launched Claude Code auto mode on March 24, 2026. Auto mode uses a safety classifier that reviews each action before execution and blocks anything that looks like it exceeds your task scope. It's a significant step forward for unattended Claude Code sessions.
However, auto mode and ClawTab auto-yes solve the same problem from different angles. Here's how they compare:
| Feature | Claude Code Auto Mode | ClawTab Auto-Yes |
|---|---|---|
| Approval method | AI safety classifier pre-screens actions | Accepts all yes/no prompts, prefers session-scoped |
| Granularity | Per-session (all or nothing) | Per-pane (different policy per agent) |
| Remote control | No (terminal only) | Yes (phone, web, Telegram) |
| Toggle without restarting | Shift+Tab to cycle modes | One tap from any device |
| Plan requirement | Team plan (Enterprise coming) | Free (open source) |
| Model requirement | Sonnet 4.6 or Opus 4.6 only | Any Claude Code version |
| Works with cron/scheduled jobs | No (requires active session) | Yes (integrated with ClawTab scheduler) |
| Multi-agent swarms | Per-session only | Independent per-pane across all agents |
Auto mode is ideal when you're working interactively in a single Claude Code session and want the AI to handle routine approvals without losing safety. ClawTab auto-yes is designed for background automation: scheduled jobs, agent swarms, overnight runs, and remote control from your phone.
You can use both together. Run Claude Code in auto mode for its safety classifier benefits, and use ClawTab on top for remote monitoring, push notifications, and the ability to toggle auto-yes on individual agents from your phone when the classifier blocks something you want to allow.
Safety Properties
Auto-yes is designed with a few deliberate constraints:
- Per-pane scope - auto-yes is enabled per tmux pane, not globally. You can auto-approve one agent while keeping manual approval on another.
- Session-scoped permissions preferred - when both "Yes" and "Yes, during this session" are available, ClawTab picks the session-scoped option. This limits the permission grant to the current Claude Code session rather than permanently.
- Only accepts yes/no prompts - auto-yes looks for options that start with "Yes". It won't auto-answer freeform questions, multiple-choice selections, or anything that doesn't match a known yes pattern.
- Explicit opt-in - you must tap the button and confirm the dialog. It never enables itself.
- Visible state - the banner is always visible when auto-yes is active, on both mobile and desktop.
- Instant disable - one tap to turn it off, effective immediately.
Auto-yes is a power-user feature. It trades manual approval for uninterrupted agent progress. Use it when you trust the task - a well-scoped refactoring job, a migration you've tested, a code review pass. Disable it when precision matters.
When to Use Auto-Yes
Batch migrations. You're running agents across 15 repositories to update an import path. Each agent will ask to write files dozens of times. Enable auto-yes and let them run.
Overnight jobs. You kick off a big refactoring agent before bed. Enable auto-yes so it doesn't stall at 2am waiting for you to approve a file write.
Multi-agent swarms. You have 6 agents running in parallel across tmux panes. Manually approving prompts across all of them is a full-time job. Enable auto-yes on the ones doing well-defined work, keep manual approval on the exploratory ones.
CI-like workflows. You're using cron jobs to run recurring agents. Auto-yes lets them complete end-to-end without human intervention.
Toggling Auto-Yes From tmux
If you live in the terminal, the ClawTab tmux plugin gives you a keybinding to toggle auto-yes without reaching for the GUI or your phone.
Press prefix + y to toggle auto-yes on the current Claude Code pane. The pane border shows the state: [Y] in green when enabled, [y] in dim gray when disabled.
You can also toggle it from the popup menu (prefix + E, Shortcuts tab), which shows the current ON/OFF status next to the option.
Setup takes one line in your .tmux.conf:
run-shell /path/to/tmux-clawtab/clawtab.tmuxSee the tmux integration guide for the full feature set including session forking, secret injection, and skill search.
