Search skills/

Claude Octopus Claude skills

An escalation layer that runs tasks past up to twelve external AI providers (Codex, Copilot, Ollama, Grok, more) with a consensus gate, plus 63 skills.

nyldn/claude-octopus 120 skills Planning, TDD & debuggingSkill & agent tooling
Our verdict Recommended with caveats

An ambitious, actively maintained multi-provider orchestration layer with a real consensus mechanism, but it is large, needs multiple external accounts or local models to use fully, and Windows support goes through WSL only.

4/5 Depth expertise shipped Note Security Review before installing 0d Maintenance Updated today 4.3K Adoption stars OK License MIT, commercial OK

What Claude Octopus does

Claude Octopus stays dormant for ordinary requests and only activates on explicit /octo:* commands, so Claude's own /init, /review and /security-review keep working as before. When invoked, it can route a task to up to twelve external providers (Codex, Antigravity CLI, Copilot, Qwen, Ollama, Perplexity, OpenRouter, OrcaRouter, OpenCode, Cursor CLI, Grok, Kimi Code), each detected only if already configured, and flags disagreement with a stated 75% consensus gate before the result reaches production. Five of the twelve providers reuse an existing subscription or local auth (Codex, Antigravity, Copilot, Cursor CLI, plus Ollama running free locally), so a full setup does not require new paid accounts.

Beyond provider orchestration, the pack ships 31 personas, 54 slash commands and 63 skills, including a Double Diamond flow (flow-discover, flow-define, flow-develop, flow-deliver) with enforced execution contracts and state tracked in .octo/STATE.md, a 'Dark Factory' mode that runs research through delivery from a spec with review only at the end, and a design-and-architecture extraction skill (skill-extract) for reverse-engineering tokens and API surfaces from a live product. It integrates with separate memory projects (claude-mem, agentmemory, deja-vu) for cross-session persistence. Needs Claude Code v2.1.14+; Linux and macOS run natively, Windows needs WSL.

Good for

  • Getting a second (or twelfth) opinion on research, architecture or security review before shipping
  • Teams who already have several AI CLI subscriptions and want them cross-checking each other
  • Structured Double Diamond projects that need phase gates and state tracking

Not for

  • Users who only want Claude's native commands with no added complexity
  • Windows users without WSL
  • Anyone not already holding several provider accounts or local models

What's inside depth 4/5

120 skill-folder entries (63 skills plus personas and flow modules) averaging 11.7 KB, with state-tracking scripts, 11 eval files and enforced execution contracts; no dedicated test suite beyond the eval set.

Skills
120
Reference docs
0
Scripts
0
Tests / evals
11
Templates
0
Avg SKILL.md
11.5 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • Claude Code v2.1.14+
  • Up to twelve optional external provider accounts or local models (Codex, Copilot, Qwen, Ollama, Perplexity, OpenRouter, OrcaRouter, OpenCode, Cursor CLI, Grok, Kimi Code, Antigravity CLI)
  • WSL for Windows use
  • Optional claude-mem / agentmemory / deja-vu for persistent memory
  • Telemetry: none found from the project itself; using an external provider sends your task data to that provider under its own terms (e.g. OPENAI_API_KEY, PERPLEXITY_API_KEY configured locally).

Security review

Review before installing

A multi-provider orchestration plugin (Double Diamond workflows, debates, parallel work packages) delivered mostly as SKILL.md files. When you invoke its commands it can send your prompts and code to external model providers you have configured, launch headless claude -p workers in git worktrees, and offer merge or PR steps after you choose. The hooks, CLI and scripts the plugin relies on were not available to us, so those parts are unreviewed.

Worth knowing

  • Makes network requestsWorkflows send your prompts and code context to whichever external providers you configure (Codex/OpenAI, Perplexity, OpenRouter, Qwen, Copilot, Cursor, etc.) using your own keys or logins; the legacy auto-router (OCTOPUS_AUTO_ROUTER_MODE=invoke) is opt-in and would share routed prompts with them.README.md:155
  • Relies on a CLI or package we could not reviewSkills dispatch to provider CLIs (codex, qwen, agy, gh, claude -p) and to the octopus helper CLI and plugin scripts, none of whose code is among the files we have..claude/skills/skill-doctor/SKILL.md:264
  • Installs packages or toolsSetup docs run npm install -g @openai/codex @qwen-code/qwen-code (unpinned) and, for Cursor, npm install in an MCP server folder.README.md:407

What we checked

The repository (1419 files) on Oct 29, 2026. Not a full review: 1376 of 1419 files were not downloaded (large repository; we reviewed the skill folders); it relies on code outside the repository: Plugin manifests, hooks (SessionStart install-metadata refresh, optional PostToolUse and context hooks), the `octopus` CLI and all scripts and agents: 1376 readable files in the repo were not available to us, so only README.md and 40 SKILL.md files were reviewed; Provider CLIs and services (Codex, Qwen, Copilot, Perplexity, OpenRouter, Cursor CLI, Grok, Kimi, Antigravity); npm packages @openai/codex and @qwen-code/qwen-code.

NVIDIA SkillSpector static scan: 37 pattern hits (Excessive Agency 14, Agent Snooping 5, Data Exfiltration 4, Tool Misuse 4, Privilege Escalation 2, Memory Poisoning 2). The YARA 'hidden instructions' and 'disable safety' hits match the sentence saying context profiles never disable safety hooks; credential hits are example text in flow-define about login features; 'rm -rf' hits are uninstall instructions; 'claude -p' is flow-parallel launching headless worker agents in git worktrees with provider API keys explicitly unset; git merge/push commands in skill-finish-branch are offered as numbered options after an AskUserQuestion prompt.

A review, not a certification. Skills change after we check them — read the files before installing.

License

MIT — commercial use allowed.

Can it run unattended?

Yes. Dark Factory mode is explicitly described as running research through delivery autonomously from a spec with review at the end, and phases track state in .octo/STATE.md so work can resume; the consensus gate adds a built-in check before anything ships.

Alternatives to Claude Octopus

Harness

revfactory/harness

A Claude Code meta-skill that designs a project-specific agent team, writes the agent definitions and skills for it, and has an evolve skill to refine them from feedback.

With caveats No issues Depth 3/5

wshobson/agents

wshobson/agents

A multi-harness plugin marketplace with 94 plugins, 202 agents, 185 skills and 105 commands for Claude Code, Codex, Cursor, OpenCode, Copilot, Antigravity and Pi.

With caveats Check first Depth 2/5

Paperclip

paperclipai/paperclip

A self-hosted app for orchestrating teams of AI agents with org charts, budgets and approvals; its skills are mostly maintainer workflows.

With caveats Broad access Depth 4/5

delegate-skills

amElnagdy/delegate-skills

Twenty skills that hand a coding task to another agent CLI (Codex, Cursor, Aider, Claude and many more) in the background, then leave the diff review and commit to you.

Recommended Broad access Depth 4/5

FAQ

What is Claude Octopus?

It is an independent, open-source escalation layer for Claude Code that routes a task to up to twelve external AI providers for a second opinion, review or debate, with a stated 75% consensus gate, on top of 63 skills and a Double Diamond workflow. It is not affiliated with Anthropic.

Do I need to pay for all twelve providers?

No. Codex, Antigravity CLI, Copilot and Cursor CLI can reuse an existing subscription or local login, and Ollama runs free locally. Only some providers like Qwen or Perplexity require their own API key.

Does it change how Claude normally behaves?

No. The README states Octopus stays dormant until you explicitly run an /octo:* command, so ordinary Claude requests and native commands such as /review are unaffected.

Does it work on Windows?

Only through WSL; the README says Linux and macOS run natively, and Windows users should run Claude Code and Octopus inside a WSL distribution rather than the desktop app's built-in WSL environment.

Sources: GitHub repository (stars, license, file tree, last push Oct 29, 2026). Review based on the README and SKILL.md files as published.