Orchestration skill for the Orca desktop app: When users ask another agent (for example, Codex within Claude) to perform a task and report back, or when they need to monitor task DAGs, ask/reply Q&A, or decision gates, the agent dispatches independent Orca workers via the Orca command line. Each worker runs in its own workspace, and the results are then collected.
skills.sh install count 353537 (+456 / 24h)
Security Audit: Not audited
- Not yet scanned
Rationale
- SKILL.md is just the entry point: Full usage is output by the orca program according to its version (ORCA skills get orchestration), avoiding inconsistencies between documentation and actual command versions
- Highlights a practical pitfall: On Linux, if you are not inside the Orca terminal, you must use orca-ide; typing orca directly usually launches GNOME's Orca screen reader and produces audio output on the user's computer
- Clear boundaries: Use the orca-cli skill when you only want to dispatch tasks without collecting results; for general long-running and parallel tasks, prioritize using the agent's own sub-agents
- Orca itself is MIT-licensed and supports macOS, Windows, and Linux. The README states it can run Codex, Claude Code, OpenCode, etc., side-by-side and offers mobile viewing; skills.sh shows 456 installs in the last 24 hours, with a cumulative total of 353537 (indicating attention level only)
Usage
- Installation command will be displayed after the audit passes; you must first install the Orca desktop app (download from onorca.dev, or run brew install --cask stablyai/orca/orca). If Orca is not running, the skill instructs the agent to execute ORCA open --json
- Orca itself is free and open-source; dispatched agents like Codex and Claude Code use their own accounts and quotas
Risks
- Not audited in this cycle, fixed at C-level, installation command not displayed; SKILL.md instructs the agent to execute orca commands and requires "rerunning with elevated permissions" when encountering runtime_access_denied, effectively requesting to bypass the sandbox. Verify this yourself before allowing it. The Gen Agent Trust Hub audit on skills.sh is marked as Warn
- The actual operational instructions are in the program output, which was not run or read in this instance; when multiple agents run in parallel, token usage and costs increase multiplicatively