Claude Octopus 🤖 Agent Open source

What it does

Claude Octopus runs a task through multiple AI models and surfaces where they disagree before you ship. Claude Code handles the ordinary path; Octopus stays dormant until you explicitly invoke an /octo:* command, then escalates to multi-model workflows: adversarial review, structured provider debates, consensus gating, and a "Dark Factory" mode that takes a spec and runs research, definition, development, and delivery autonomously. It integrates a dozen external providers (Codex, Antigravity CLI, Copilot, Qwen, Ollama, Perplexity, OpenRouter, OrcaRouter, OpenCode, Cursor CLI, Grok, and Kimi Code) alongside the built-in Claude host, but needs none of them to start.

What is inside

The plugin ships dozens of specialized personas (security-auditor, backend-architect, and the like), a large set of slash commands, and reusable skill modules, plus engineering methods for architecture, TDD, and debugging. Workflows include a Double Diamond discover-to-deliver pipeline, a multi-LLM council with goal modes and adversarial styles, parallel workstreams in isolated git worktrees, and multi-model PR review with inline comments. Optional session memory integrations persist decisions across sessions. Diagnostics (doctor, capabilities, repair, security-audit) check and repair the installation offline.

Works with

Claude Code and Codex as plugins, Cursor as an MCP server, plus OpenCode and Factory Droid. External providers are added one at a time and run only inside explicit workflows.

How to install

claude plugin marketplace add https://github.com/nyldn/plugins.git
claude plugin install octo@nyldn-plugins

Maintenance and safety

MIT licensed and very actively developed. Nothing routes automatically: ordinary prompts are untouched unless you opt into a suggest or invoke router, and the invoke mode can start paid external-provider workflows and share prompt context with those providers, so provider-side data retention follows each account's policy. It is an independent project, not affiliated with Anthropic.

Who should use it

Developers who want cross-model second opinions and adversarial review on high-stakes work, and who are comfortable managing multiple provider credentials.