AI Auto-Accept

Stop clicking "Yes". AI Auto-Accept answers the permission questions AI coding agents ask ("Allow this bash command?", "Do you want to make this edit?") for you: in the Claude Code chat panel, and in AI CLIs running in VS Code terminals (Claude Code, Codex, Gemini, Amazon Q, Aider). Dangerous commands (rm -rf, git push --force, DROP TABLE, terraform destroy, ...) are never answered: those questions are always left to you.

Quick start

Claude Code chat panel

1. Install AI Auto-Accept and the Claude Code extension.
2. Open the Command Palette (Cmd+Shift+P on macOS, Ctrl+Shift+P on Windows/Linux), run AI Auto-Accept: Set Up Chat Auto-Approve and pick Recommended. To set up only the Claude Code part, add "claudeAutoAccept.claudePanel": true to your User settings instead.
3. Use Claude as usual. When Claude is about to ask for permission, the question is answered Yes for you, so the command or edit just runs.

This works in every Claude chat at the same time: the side bar, chats opened in editor tabs, tabs in the background, and chats that were already open when you turned it on. Each chat runs its own Claude process, and every one of them uses the same hook.

Claude asks most in Manual mode (before almost every edit and command), so that is where you see the most answered for you. In Auto mode, Claude's own safety check decides and rarely asks.

The answer comes from a hook saved in ~/.claude/settings.json, so it also answers the claude CLI in any terminal, even while VS Code is closed. It keeps the on/off state you last set in VS Code.

AI CLIs in a terminal

Nothing to set up. Run claude, codex, gemini, q or aider in a VS Code terminal, and their permission prompts are answered for you. The terminal needs shell integration, which is on by default for bash, zsh, fish and PowerShell.

How to tell it's working

• The status bar (bottom right) shows ✓ Auto-Accept 3. The number is how many questions were answered for you since the window opened.
• Run AI Auto-Accept: Show Log to see each one, for example ACCEPT [Claude Code] "Bash: npm test".
• A dangerous command still asks you, and the log says why: BLOCKED [Claude Code] "Bash: rm -rf build" (deny-list: answer it yourself).

Turn it off

• Click Auto-Accept in the status bar, or run AI Auto-Accept: Toggle. While it is off, nothing is answered for you, in the Claude chat or in terminals.
• To remove the Claude Code hook completely, run AI Auto-Accept: Set Up Chat Auto-Approve and pick Undo, or set claudeAutoAccept.claudePanel to false.

Troubleshooting

Upgrading from 0.1.6

0.1.6 swapped the Claude Code chat support for a "chat auto-click" setting that could not reach the Claude Code panel. From 0.1.8 the hook is back. If you turned on chat auto-click in 0.1.6, run AI Auto-Accept: Set Up Chat Auto-Approve → Recommended again, or set claudeAutoAccept.claudePanel to true. You can delete the leftover claudeAutoAccept.chatPanel setting.

Supported agents (terminal)

The extension only acts on prompts it recognises. Ordinary output that happens to contain a question ("Do you want to know more?") is ignored.

Chat panels (GitHub Copilot agent mode, the Claude Code VS Code panel) are webviews, not terminals. See Chat panels.

How it works

1. VS Code's shell integration API (onDidStartTerminalShellExecution) streams the output of every command you run in a terminal, such as claude or codex.
2. The extension keeps the last ~4000 characters with ANSI escape codes removed. Once the output has been quiet for about 200 ms, it checks the text for a known prompt.
3. If it finds one and the prompt text doesn't match the deny-list, it sends the keys shown above to the terminal. The status bar counter goes up by one.
4. If the same prompt keeps coming back (a stuck prompt), the extension answers it at most maxRepeat times (3 by default), then stays quiet until a different prompt appears or that prompt has been gone for 10 s.

Every decision (accepted, blocked, gave up, dry run) is written to the AI Auto-Accept output channel. Run AI Auto-Accept: Show Log to open it.

Safety

• User settings only. denyList, agentCommands and promptPatterns are read from your user settings only. A repository's .vscode/settings.json cannot switch off the deny-list or widen what gets answered.
• Deny-list (on by default). A prompt whose text matches any claudeAutoAccept.denyList regex is never answered. The patterns are also checked against the command box with its wrapped lines rejoined, so a long command the agent wraps (git push origin my-branch on one line, --force on the next) is still caught. The log records which pattern matched, and you answer the prompt yourself. The default list covers recursive rm (-r, -R, -rf, --recursive, flags in any order), git push --force/+refspec, find -delete, curl … | sh/python/node, git reset --hard, git clean -f, dd of=, destructive database commands (DROP, TRUNCATE, DELETE FROM, UPDATE … SET without WHERE, ALTER TABLE … DROP, dropdb, Mongo drop()/deleteMany({}), Redis FLUSHALL, prisma migrate reset, rails db:drop, artisan migrate:fresh, manage.py flush, supabase db reset), infrastructure teardown (terraform destroy, kubectl delete, helm uninstall, aws s3 rm --recursive, aws … delete-*, docker volume prune), mkfs, chmod 777, writes to raw disks, and fork bombs. The deny-list is a safety net, not a sandbox: a command written in a way the patterns don't cover will still be accepted.
• Dry run. Set claudeAutoAccept.dryRun to log what would be sent without sending anything.
• Per-agent switches. You can turn off any agent under claudeAutoAccept.agents.
• Toggle. Click the status bar item, or use AI Auto-Accept: Toggle.

Chat panels

The VS Code chat (GitHub Copilot agent mode) and the Claude Code panel are not terminals, so the extension can't read their prompts. Run AI Auto-Accept: Set Up Chat Auto-Approve instead and pick a level:

Recommended only turns on the terminal-tool rules and the Claude Code hook; it does not turn on claudeAutoAccept.copilotPanel (below), since VS Code's own terminal auto-approve already covers most Copilot chat tool calls. Turn that setting on yourself for full tool-level coverage (file edits, MCP tools, ...) in Copilot chat too.

How the hooks work (Claude Code panel/CLI: PermissionRequest; VS Code chat/Copilot: PreToolUse):

• They work in already-open sessions, not just new ones.
• They follow the status-bar on/off switch, agents.claude (Claude) and dry run.
• Each Claude Code question answered Yes adds one to the status bar counter and is logged as ACCEPT [Claude Code] … in the AI Auto-Accept output channel. Deny-listed ones are logged as BLOCKED and left to you.
• They run with VS Code's built-in Node runtime, so Node doesn't need to be installed. On Windows, node must be on PATH.
• The Claude hook respects CLAUDE_CONFIG_DIR if you've set it; the Copilot hook respects COPILOT_HOME.

After choosing a level, start a new VS Code chat session, because open chats keep their approval mode (except under "Everything"). If your organisation's policy disables auto-approve, VS Code ignores these settings.

Claude Code panel opt-in (without "Set Up Chat Auto-Approve")

Set claudeAutoAccept.claudePanel to true directly to install just the PermissionRequest hook described above, without touching any chat.* settings.

VS Code chat / Copilot panel opt-in

Set claudeAutoAccept.copilotPanel to true to install a PreToolUse hook at ~/.copilot/hooks/ai-auto-accept.json. It approves every tool call in VS Code chat / GitHub Copilot agent mode (including file edits and MCP tools, not just terminal commands) except deny-listed ones, which are left for you to answer.

Settings

{
  "claudeAutoAccept.enabled": true,
  "claudeAutoAccept.agents": { "claude": true, "codex": true, "gemini": true, "amazonq": true, "aider": true, "generic": true },
  "claudeAutoAccept.denyList": [ /* built-in list, see Safety */ ],
  "claudeAutoAccept.agentCommands": "^(claude(-code)?|codex|gemini(-cli)?|q|qchat|aider|copilot|opencode)$",
  "claudeAutoAccept.promptPatterns": [],   // extra regexes, answered with y + Enter
  "claudeAutoAccept.dryRun": false,
  "claudeAutoAccept.maxRepeat": 3,
  "claudeAutoAccept.claudePanel": false,   // opt-in; PermissionRequest hook, deny-list still applies
  "claudeAutoAccept.copilotPanel": false   // opt-in; PreToolUse hook, deny-list still applies
}

The extension logs and skips any invalid regex in denyList or promptPatterns. If you set denyList yourself, your list replaces the default, so copy the default entries into it if you want to keep them.

Commands

• AI Auto-Accept: Enable / Disable / Toggle (the legacy claude-auto-accept.* commands still work)
• AI Auto-Accept: Show Log
• AI Auto-Accept: Set Up Chat Auto-Approve

Limitations

• Only AI agent commands are watched. A terminal command is watched only if it matches claudeAutoAccept.agentCommands (the program name, by default claude, claude-code, codex, gemini, q, qchat, aider, copilot, opencode; npx/node launchers are looked through). A [Y/n] from apt, npm or prisma is never answered. Prompts whose default is No ([y/N]) are never answered, even from agents.
• Requires shell integration. This is on by default for bash, zsh, fish and PowerShell. If a terminal has no shell integration after 10 s, the extension logs a warning and adds it to the status bar tooltip. That terminal is not watched.
• Commands that were already running before the extension activated (for example, after a window reload) are missed. Restart the agent to fix this.
• Detection is pattern-based. A new CLI version that rewords its prompts may stop being recognised until the patterns are updated. Use promptPatterns in the meantime.
• Chat panels are covered through their own approval settings and the Claude Code hook, described in Chat panels.
• Requires VS Code 1.93 or later.

Development

npm install
npm run test:unit  # compiles, then runs the detector and hook unit tests
npm run test:e2e   # launches real VS Code against test/fake-agent.js scenarios