🐙 Claude Octopus

Every AI model has blind spots. Claude Octopus supports twelve external provider integrations — Codex, Antigravity CLI, Copilot, Qwen, Ollama, Perplexity, OpenRouter, OrcaRouter, OpenCode, Cursor CLI, Grok, and Kimi Code — alongside the built-in Claude Code host, with consensus gates that flag disagreements before you ship.

Claude-native first, Octopus for escalation. Use Claude-native /init, /review, and /security-review when Claude is enough. Use Octopus when you want multiple model opinions, adversarial review, or stricter multi-LLM workflows.

Claude Octopus is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Anthropic.

🐙 Research, build, review, and ship — with twelve external providers checking the host's work. Claude-native handles the ordinary path. Octopus remains dormant until you explicitly run /octo:*, then handles the escalated path. A 75% consensus gate catches disagreements before they reach production.

🧠 Remembers across sessions. Integrates with claude-mem, agentmemory, and deja-vu for persistent memory — past decisions, research, and context survive session boundaries.

⚡ Spec in, software out. Dark Factory mode takes a spec and autonomously runs the full pipeline — research, define, develop, deliver. You review the output, not every step.

🔄 Choose the workflow the task needs. Use a focused method for architecture, debugging, or TDD. Use /octo:embrace for Discover → Define → Develop → Deliver, with quality gates between phases.

🐙 31 specialized personas (role-specific AI agents like security-auditor, backend-architect), 54 commands (slash commands you type), 63 skills (reusable workflow modules). Explicit workflows select the experts they need; ordinary Claude requests do not activate Octopus.

🐙 Works with just Claude. Adds up to twelve external provider integrations. Zero external providers are needed to start. Add them one at a time — each becomes available when detected and runs only inside an explicit workflow.

💰 Five providers cost nothing extra when you already have access. Codex, Antigravity CLI, Copilot, and Cursor CLI use existing subscriptions or local auth. Ollama runs locally for free. Qwen now requires API-key or Coding-Plan auth; its free OAuth tier ended on 2026-04-15.

What's New

Reliable fallback routing and host detection

Research, design, copywriting, and image tasks can use authenticated Codex or installed Claude seats when Antigravity is unavailable. Provider configuration loads before fallback selection.

Claude Code sessions also identify their active host correctly when the shared plugin is installed in a Codex cache. This keeps Codex council seats independent. See provider availability and host detection.

Safer model configuration and reliable provider history

Model configuration rejects unsafe names while preserving supported Antigravity display labels. Concurrent provider-history updates retain entries even when a directory utility reports success to multiple writers. Optional history recording requires Python 3 and skips the write if it cannot acquire the lock.

The MCP SDK is updated to 1.31.0, including an upstream OAuth client security fix. See the changelog for details.

Parallel work, reliable reviews and Codex startup

Parallel work packages now use separate branches. Completed commits remain available after cleanup, and worktrees with uncommitted edits are kept for recovery. Failed packages record their completion status so the wave can finish.

Reviews and research retain provider stdout longer than the configured threshold, even when it contains a context-limit rejection phrase. Short stdout-only rejections and stderr rejection signatures still fail the affected seat. The new OCTO_PROVIDER_REJECTION_MAX_OUTPUT_BYTES setting adjusts the output-size threshold, with a default of 4096 bytes. See provider rejection handling.

Codex can create its thread-coordination and plugin-sync locks inside the Linux Tangle execution boundary. The locks use private temporary storage, while configuration and extension inputs remain read-only. See bounded Codex runs for host requirements and temporary-directory configuration.

Engineering methods

Octopus includes eight engineering methods adapted from Matt Pocock's skills. Routine architecture, TDD, and debugging use your current host. Ask for an independent opinion when a reviewer would help. Plans capture domain terms and blocking decisions, compare interface designs, and can propose a time-limited prototype.

Setup can resume an interrupted configuration and rechecks readiness before reporting success. See workflow methods for usage and the changelog for release details.

Premium /octo:auto routes also run one bounded cross-provider peer check after an eligible single-owner result, without requiring a second command or flag. Budget and Standard routes do not add the check, and existing multi-model workflows are not double-reviewed. Set OCTOPUS_PREMIUM_PEER_CHECK=off to disable it.

🆕 v11.14.0 — LiteLLM gateways, fewer Tangle review retries, and Command Code runs that stop when canceled.

Default roster: Claude Opus 5.5 leads architecture, planning, security reasoning, and final judgment; GPT-5.6 Sol is the independent implementation/review peer; Claude Sonnet 5.5 is the standard Claude seat; Fable 5.1 remains an opt-in judgment escalation. Existing model pins and provider configuration still win. See the routing strategy.

/octo:model-config                         # inspect or override the frontier roster
OCTOPUS_OPUS5_AUTO_XHIGH=1                 # opt in to automatic xhigh Opus 5 phases
OCTOPUS_OPUS_MODEL=claude-fable-5-1        # explicitly opt in to Fable 5.1
OCTOPUS_CODEX_MODEL=gpt-6-astra            # explicitly opt in to Astra
/octo:model-config tier premium claude claude-fable-5-1  # one bounded Fable judgment seat
/octo:model-config tier premium codex gpt-6-astra        # one bounded Astra judgment seat

🆕 v9.41 — Multi-LLM Council. /octo:council runs a structured 3/5/7-persona deliberation across Claude, Codex, Antigravity, and OpenCode with goal modes (advice, decision, plan, implement, review), styles (balanced, adversarial, red-team, executive, implementation), benchmark-aware role routing, quorum + critical-veto gates, budget caps, and gated worktree handoff for approved plans. Use it when one model's opinion isn't enough.

/octo:council --goal decision --style adversarial "Should this service stay monolithic?"
/octo:council --goal implement --implement plan-only "Refactor the auth flow"