---
title: "Orca YOLO Mode: The Agent Bypass Flags and Why They Need a Sandbox"
date: "2026-09-19"
video_id: "7687282408898432286"
tiktok_url: "https://www.tiktok.com/@imjonathanacuna/video/7687282408898432286"
creator: "@imjonathanacuna"
category: "AI"
tags: ["orca", "claude code", "codex", "yolo mode", "ai coding agents"]
excerpt: "Orca launches Claude Code, Codex, Grok and Gemini with permission bypass on by default. What each flag does, and why every vendor says to use a container."
---
Orca is a desktop workspace that runs several AI coding agents (Claude Code, Codex, Grok, Gemini and others) side by side, each in its own git worktree. It launches every supported agent in "Yolo" mode by default, which passes each CLI's permission-bypass flag. Every vendor that documents those flags says to run them inside a container or VM, and a worktree is neither.

A TikTok from @imjonathanacuna calls Orca "the most dangerous number one AI tool known to mankind" and walks through Settings, then AI agents, to confirm each agent launches in Yolo mode. The flags he shows are real. The setup step is already the default, and the video runs everything on a personal laptop.

## What is Orca?

Orca is an MIT-licensed app from [stablyai](https://github.com/stablyai/orca) for running multiple coding agents in parallel. Its Agents settings page offers Claude, Claude Agent Teams, Codex, Grok, OpenCode, Gemini and Hermes as default agents. Orca also ships a mobile companion on iOS and Android for monitoring and steering agents, per the project README.

Orca ADE [announced Claude Agent Teams as a built-in terminal mode](https://x.com/orca_build/status/2074579289757139057), with a dedicated pane showing the whole agent team. Its command in the video is `orca claude-teams` with `--dangerously-skip-permissions` in the Arguments field.

## What does Orca's Yolo mode do?

Orca's [supported agents documentation](https://www.onorca.dev/docs/agents/supported) says Orca pre-fills each supported CLI's permission-bypass flag for new launches. It names `--dangerously-skip-permissions` for Claude, `--dangerously-bypass-approvals-and-sandbox` for Codex, and `--yolo` for Gemini, Cursor, Crush, Kimi, Rovo Dev, Hermes, GitHub Copilot and Command Code.

The [settings reference](https://www.onorca.dev/docs/settings) describes the Agent Permissions control as choosing "Yolo for fewer CLI permission prompts, or Manual to keep each agent's own approval flow for agents you haven't customized."

Yolo is the shipped state, and Manual is the opt-out. A fresh install behaves the way the video describes without opening settings. The walkthrough is still useful for checking which agents picked up a flag. In the video, OpenCode shows a bare `opencode` command, and Hermes shows a bare `hermes` with its Arguments field cut off.

## The four bypass flags, from each vendor's docs

**Claude Code.** `claude --dangerously-skip-permissions` is defined in Anthropic's [CLI reference](https://code.claude.com/docs/en/cli-reference) as "Skip permission prompts. Equivalent to `--permission-mode bypassPermissions`." The [permission modes page](https://code.claude.com/docs/en/permission-modes) lists the recommended use for `bypassPermissions` as "Isolated containers and VMs only." For unattended runs, the same page marks a container, VM or the sandbox runtime as "Required," running as a non-root user on Linux and macOS.

**Codex.** `codex --dangerously-bypass-approvals-and-sandbox` appears in OpenAI's [developer commands reference](https://learn.chatgpt.com/docs/developer-commands?surface=cli) with `--yolo` as an alias. OpenAI describes it as "Run every command without approvals or sandboxing. Only use inside an externally hardened environment."

**Grok.** `grok --permission-mode bypassPermissions` is the long form of `--always-approve` (alias `--yolo`) in xAI's [permissions and safety guide](https://github.com/xai-org/grok-build/blob/main/crates/codegen/xai-grok-pager/docs/user-guide/22-permissions-and-safety.md). Deny rules, hooks and some shell ask rules still apply, and an administrator can turn the mode off with `disable_bypass_permissions_mode = true`.

**Gemini CLI.** `gemini --yolo` is defined in Google's [configuration docs](https://google-gemini.github.io/gemini-cli/docs/get-started/configuration.html) as "Enables YOLO mode, which automatically approves all tool calls."

## Is a git worktree a sandbox?

Orca's case for defaulting to bypass is that each agent works in its own git worktree, so a bad diff can be thrown away. [SFEIR's write-up of Orca](https://www.sfeir.com/articles/orca-ade-flotte-agents-paralleles/) points out the limit. A worktree isolates versioned files. The agent still runs as the user, with access to SSH keys, cloud tokens, `.env` files, the network and every other repository on the disk. Containment needs a container, a micro-VM, or a machine whose compromise does not matter.

Anthropic says containers and VMs only. OpenAI says an externally hardened environment only. The video runs all of it on what appears to be the presenter's everyday MacBook, judging by the personal home directory path on screen. The "dangerous" label in the hook is accurate, and the video gives no advice on reducing the risk.

## Does Claude Code's auto mode change the tradeoff?

Anthropic's permission modes page says Pro, Max and Team plans now start in auto mode, where a classifier reviews actions instead of the user. Auto mode already removes many prompts, so full bypass buys less convenience than it used to while carrying the same exposure.

## Can Orca run 50 agents at once?

The video says "we're trying to spawn 50 agents at the same time to work on tons of parallel projects all from one mobile app or one app." I found no published concurrency limit for Orca to check that against, so treat 50 as a goal.

Subscription limits are the more likely ceiling. The presenter's status bar shows one account at `100% used` with about thirteen hours until reset. The Claude Code header in the video reads `Haiku 4.5 · Claude Max`, and Max plans run a five-hour rolling window alongside a weekly cap. Fifty agents on a few consumer subscriptions will hit those meters first.

## Bottom line

- The four bypass commands in the video are real and documented, and Orca turns them on by default.
- Anthropic, OpenAI and xAI all document these flags for containers, VMs or hardened environments.
- A worktree protects your repo history. It does not protect your keys, tokens or other projects.

## Sources

- [stablyai/orca on GitHub](https://github.com/stablyai/orca): Orca source, agent list and mobile apps.
- [Orca supported agents](https://www.onorca.dev/docs/agents/supported): the default bypass flag for each CLI.
- [Orca settings reference](https://www.onorca.dev/docs/settings): the Yolo / Manual control wording.
- [Claude Code CLI reference](https://code.claude.com/docs/en/cli-reference): `--dangerously-skip-permissions`.
- [Claude Code permission modes](https://code.claude.com/docs/en/permission-modes): `bypassPermissions` and auto mode.
- [OpenAI Codex developer commands](https://learn.chatgpt.com/docs/developer-commands?surface=cli): `--dangerously-bypass-approvals-and-sandbox`.
- [xAI Grok permissions and safety](https://github.com/xai-org/grok-build/blob/main/crates/codegen/xai-grok-pager/docs/user-guide/22-permissions-and-safety.md): `bypassPermissions` and the admin lock.
- [Gemini CLI configuration](https://google-gemini.github.io/gemini-cli/docs/get-started/configuration.html): `--yolo`.
- [SFEIR on Orca](https://www.sfeir.com/articles/orca-ade-flotte-agents-paralleles/): why a worktree is not a sandbox.
- [Orca ADE on X](https://x.com/orca_build/status/2074579289757139057): the Claude Agent Teams announcement.


## Video description and verification notes

**The four launch flags shown in this video are real, documented, and every vendor that publishes them says to run them inside a container or VM, which is not where the video runs them.** This is a 70-second vertical TikTok from Jonathan Acuña, who posts as "Doctor AI" (@imjonathanacuna). It opens with a talking-head shot, a persistent white "How to Setup Orca" title card pinned near the top, the Orca logo above it, and burned-in single-word captions at the bottom (DANGEROUS, GO, PERMISSIONLESS, SUBSCRIPTIONS). After the first few seconds it cuts to a handheld phone shot pointed at a MacBook screen, with the presenter's ringed hand moving across the trackpad and a green highlight box drawn in post to mark each click. The claim is that Orca is "the most dangerous number one AI tool known to mankind," and the walkthrough covers exactly one settings page: Settings, then AI agents, then confirming every agent launches in YOLO mode.

## What is actually on the screen

The frames give more detail than the narration does. The terminal pane in the opening desktop shot reads `Claude Code v2.1.274`, `Haiku 4.5 · Claude Max`, and a working directory of `~/orca/workspaces/VS Code Programs/Orca-Demo`. The status line below it reads `jjacuna/Orca-Demo`, a worktree ID beginning `wt 79626-88752204-9f58-4ce0-9456-c912008`, and `bypass permissions on (shift+tab to cycle) · 3 agents`. A Claude Code SessionStart hook has fired a message about the Warp plugin being installed while the session is not running in Warp.

The General settings page shows a Workspace Directory of `/Users/jonathanacuna/orca/workspaces`, a Sources list containing `Claude Code · .claude/worktrees/*` and `GSD · .gsd-workspaces/*`, a Worktree root field, and a Nest Workspaces option described as "Create workspaces inside a repo-named subfolder."

The Agents page is where the walkthrough lands. The Default Agent row offers Auto, No agent (blank terminal), Claude, Claude Agent Teams, Codex, Grok, OpenCode, Gemini, and Hermes, with Claude selected. Below it sit Agent status hooks, Auto-generate tab titles, Keep computer awake, a Prompt Cache Timer, and then Agent Permissions with a Yolo / Manual toggle. The Installed list reads `7 detected` and shows each agent with its launch command underneath.

## The four flags, checked against their own documentation

Each command string visible in the Installed list resolves to a real, documented flag.

`claude --dangerously-skip-permissions` appears under Claude. Anthropic's [CLI reference](https://code.claude.com/docs/en/cli-reference) defines it as "Skip permission prompts. Equivalent to `--permission-mode bypassPermissions`." The [permission modes page](https://code.claude.com/docs/en/permission-modes) lists `bypassPermissions` as running "Everything," with the recommended use case stated as "Isolated containers and VMs only." The same page carries a table row for running unattended that lists a container, VM, or the sandbox runtime as "Required," and specifies running as a non-root user on Linux and macOS.

`codex --dangerously-bypass-approvals-and-sandbox` appears under Codex. OpenAI's [developer commands reference](https://learn.chatgpt.com/docs/developer-commands?surface=cli) lists the flag with `--yolo` as an alias and describes it as "Run every command without approvals or sandboxing. Only use inside an externally hardened environment."

`grok --permission-mode bypassPermissions` appears under Grok. xAI's [permissions and safety guide](https://github.com/xai-org/grok-build/blob/main/crates/codegen/xai-grok-pager/docs/user-guide/22-permissions-and-safety.md) documents `--always-approve` with `--yolo` as an alias and `--permission-mode bypassPermissions` as the equivalent long form. That guide notes deny rules, hooks, and some shell ask rules still apply, and that an administrator can disable the mode entirely with `disable_bypass_permissions_mode = true`.

`gemini --yolo` appears under Gemini. Google's [Gemini CLI configuration docs](https://google-gemini.github.io/gemini-cli/docs/get-started/configuration.html) define `--yolo` as "Enables YOLO mode, which automatically approves all tool calls."

Claude Agent Teams is the one entry that is Orca-specific. Its command is `orca claude-teams` with `--dangerously-skip-permissions` in the Arguments field. Orca ADE [announced Claude Agent Teams as a built-in terminal mode](https://x.com/orca_build/status/2074579289757139057), with a dedicated pane showing the whole agent team.

## The setup step the video demonstrates is already the default

The narration frames this as configuration work: "you're pretty much just going to be matching the settings that I have," and "just make sure it's enabled for these." Orca's own [supported agents documentation](https://www.onorca.dev/docs/agents/supported) states that Orca pre-fills each supported CLI's permission-bypass flag for new launches, naming `--dangerously-skip-permissions` for Claude, `--dangerously-bypass-approvals-and-sandbox` for Codex, and `--yolo` for Gemini, Cursor, Crush, Kimi, Rovo Dev, Hermes, GitHub Copilot, and Command Code. The [settings reference](https://www.onorca.dev/docs/settings) describes the Agent Permissions control as choosing "Yolo for fewer CLI permission prompts, or Manual to keep each agent's own approval flow for agents you haven't customized."

That means the Yolo toggle the video points at is the shipped state, and Manual is the opt-out. A fresh Orca install behaves the way the video describes without touching the gear icon. What the walkthrough is useful for is confirming nothing got flipped and seeing which of the seven detected agents actually picked up a flag. Two entries in the frames have no bypass argument visible: OpenCode shows a bare `opencode`, and Hermes shows a bare `hermes` with its Arguments field cut off below the frame edge.

## A worktree isolates files, not the machine

Orca's design case for defaulting to bypass is that each agent runs in its own git worktree, so a bad diff is discardable. [SFEIR's write-up of Orca](https://www.sfeir.com/articles/orca-ade-flotte-agents-paralleles/) takes that argument apart directly: a worktree isolates versioned files, not a system. The agent keeps the operator's user account, SSH keys, cloud tokens, `.env` files, network access, and every other repository on the disk. Real containment needs a container, a micro-VM, or a machine whose compromise does not matter.

That is the same boundary every vendor above named. Anthropic says containers and VMs only. OpenAI says externally hardened environment only. The video runs all of it on what appears to be the presenter's daily-driver MacBook, evidenced by the personal home directory path and the Warp plugin hook message. The "dangerous" framing in the hook turns out to be accurate rather than promotional, and the video does not return to it with any mitigation advice.

One more piece of context makes the tradeoff narrower than it looks. Anthropic's permission modes page states that on Pro, Max, and Team plans the built-in starting mode is now auto mode, where a classifier reviews actions instead of the user. The gap between auto mode and full bypass is smaller than the gap between manual approval and full bypass, so the convenience won by the flag has shrunk while the exposure has not.

## The 50-agent number and the subscription math

The narration says "we're trying to spawn 50 agents at the same time to work on tons of parallel projects all from one mobile app or one app." Orca does ship a mobile companion on iOS and Android for monitoring and steering agents, per the [project README](https://github.com/stablyai/orca), so the "one app" part holds. The 50 figure is a stated ambition in the video and I found no published concurrency cap for Orca to check it against, so treat it as an aspiration rather than a measured limit.

The binding constraint is more likely the subscriptions themselves. The presenter's own status bar makes the point: across the frames it reads `0% used Fable`, `10% used 5d 2h`, and `100% used 12h 51m`. One of his accounts is fully consumed with roughly thirteen hours to go before reset. Anthropic's Max plans run a five-hour rolling window alongside a weekly cap, and Claude Code's header in this video reads `Haiku 4.5 · Claude Max`. Fifty concurrent agents on a handful of consumer subscriptions runs into those meters long before it runs into anything Orca does. The video's own framing supports this, since it describes "managing multiple subscriptions" as the whole goal.

## Transcript and caption accuracy

The Whisper transcript is close to clean and matches the burned-in captions where they overlap. Two spots drift. The auto-generated video description renders Claude as "Cloud" in the phrase "turning on Cloud, Codex, GROK," while the transcript itself gets "Claude, Codex, Grok" right. The transcript also produces "what he starts them with is basically YOLO modes," which reads as a mishearing of a reference to the tool rather than a person, and "permissionless and permission mode or permission free mode" is a garbled run of the presenter reaching for the same idea three times. Neither error changes the meaning of the walkthrough.

## Key Takeaways

- All four permission-bypass commands shown on screen are real and documented: `claude --dangerously-skip-permissions`, `codex --dangerously-bypass-approvals-and-sandbox`, `grok --permission-mode bypassPermissions`, and `gemini --yolo`.
- Anthropic, OpenAI, and xAI each document these flags with an explicit instruction to run them inside a container, VM, or hardened environment. The video runs them on a personal laptop.
- The Yolo toggle the video walks you to is Orca's shipped default, not something you enable. Manual is the opt-out.
- A git worktree isolates versioned files. It does not isolate SSH keys, cloud tokens, env files, or other repos on the same disk.
- Anthropic now starts Pro, Max, and Team sessions in auto mode with a classifier, which narrows what full bypass buys you.
- The transcript is accurate enough to follow. The auto-generated description writes "Cloud" for Claude, and the phrase "what he starts them with" is a mishearing.
- Unverified: the 50-simultaneous-agent target has no published Orca concurrency limit to check against. The Hermes entry in Orca's agent list is probably Nous Research's Hermes Agent, but Orca's docs do not name the vendor and the video frame cuts off its arguments field. The `GSD` source path in the General settings and the `Gremlin hidden` label in the status bar are not explained anywhere I could find.

## Resources

- [stablyai/orca on GitHub](https://github.com/stablyai/orca) -- the MIT-licensed source for Orca, with the agent list, mobile apps, and remote runtime described by the maintainers.
- [Orca supported agents documentation](https://www.onorca.dev/docs/agents/supported) -- the page that states Orca pre-fills each CLI's permission-bypass flag by default, naming each flag.
- [Orca settings reference](https://www.onorca.dev/docs/settings) -- the exact wording of the Agent Permissions Yolo / Manual control the video clicks.
- [Claude Code CLI reference](https://code.claude.com/docs/en/cli-reference) -- Anthropic's definition of `--dangerously-skip-permissions` and its equivalence to `--permission-mode bypassPermissions`.
- [Claude Code permission modes](https://code.claude.com/docs/en/permission-modes) -- the table listing `bypassPermissions` as for isolated containers and VMs only, plus the actions no mode auto-approves.
- [OpenAI Codex developer commands](https://learn.chatgpt.com/docs/developer-commands?surface=cli) -- the `--dangerously-bypass-approvals-and-sandbox` entry and its externally-hardened-environment warning.
- [xAI Grok permissions and safety](https://github.com/xai-org/grok-build/blob/main/crates/codegen/xai-grok-pager/docs/user-guide/22-permissions-and-safety.md) -- documents `--permission-mode bypassPermissions` and the admin lock that disables it.
- [Gemini CLI configuration](https://google-gemini.github.io/gemini-cli/docs/get-started/configuration.html) -- the `--yolo` and `--approval-mode` definitions.
- [SFEIR on Orca and its default YOLO mode](https://www.sfeir.com/articles/orca-ade-flotte-agents-paralleles/) -- the clearest published argument that a worktree is not a sandbox.
- [Jonathan Acuña on TikTok](https://www.tiktok.com/@imjonathanacuna) -- the poster's account, for context on the rest of his AI tooling content.
