skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- fe7241058b1a
- Files read
- 11
This skill pack connects AI coding agents to the Orca desktop IDE's CLI so they can manage workspaces, terminals, worker agents, browsers, mobile emulators, desktop GUI windows, and Jira or Linear tickets. Every skill instructs the agent to bypass sandbox restrictions when connecting to Orca, which can expose desktop GUI control, other terminals, and saved Jira or Linear accounts through the single orca CLI if untrusted ticket, web, or repository content tricks the agent.
What it can reach
- Shell commands
- Runs the local orca CLI, controls Orca terminal sessions, spawns worker agents in worktrees, and manages VM, SSH, or container workspace recipes.
- System
- Launches the Orca desktop app, drives visible desktop windows with mouse clicks, keystrokes, and screenshots via orca computer, and controls connected Android devices and iOS Simulators.
- Browser
- Opens, interacts with, and screenshots local or external web pages in Orca's embedded browser as well as Chrome, Edge, and Safari windows.
- Secrets
- Reads, searches, creates, and updates Jira issues and Linear tickets using the accounts saved in Orca.
- Files
- Creates and manages git worktrees, workspace folders, environment recipe configs, and Orca plugins.
- Network
- Fetches external web pages through Orca's browser, calls Jira and Linear APIs through Orca, and connects to remote sandboxes or SSH hosts.
Findings
-
MediumSkills instruct the agent to bypass its sandbox for the all-in-one Orca CLI
All eleven skills instruct the agent to automatically re-run commands with escalated (unsandboxed) permissions whenever the agent's sandbox blocks Orca's local connection, and to start the Orca desktop app if it is not running. Several skills activate automatically on everyday tasks, such as checking how an external web page renders or working in a folder with an orca-plugin.json file, and unlike the Jira and Linear skills, the browser and GUI skills do not warn the agent to treat external content as untrusted. Because reading help text, rendering web pages, reading tickets, sending commands to other terminals, using saved Jira and Linear connections, and clicking or typing in any visible desktop or browser window all go through the same orca executable, granting unsandboxed access (especially with an always-allow rule for orca) exposes all of those powers if untrusted web, repository, or ticket text tricks the agent.
Fix: Require explicit user confirmation before launching Orca or escalating out of the sandbox, advise scoping command approvals to specific subcommands (such as orca skills get or read-only queries) rather than the bare orca executable, and add untrusted-content warnings to the browser and computer-use skills.
