claude-octopus keeps twelve providers dormant until you ask for them

The premise is that every model has blind spots

The readme opens with a single claim about models in general: each one has blind spots, which is the argument for running several. The implementation of that argument is a list of twelve external provider integrations, named in the file as Codex, a coding agent CLI, Copilot, Qwen, a local model runner, a search oriented provider, two routers, OpenCode, the Cursor command line, Grok and a Kimi code seat. Alongside them sits the built in host, and the mechanism that does the work is described as consensus gates that flag disagreements before you ship. One line near the top is worth repeating because it is a legal fact rather than a feature: the project states it is independent, and is not affiliated with, endorsed by or sponsored by the company whose name appears in its title.

Nothing external runs until you type a prefixed command

The activation model is the most useful thing in the file. It is described as Claude native first with the octopus reserved for escalation: there are host commands for initialising, reviewing and security reviewing, and you use them when one model is enough. Then the framing sentence: the octopus remains dormant until you explicitly run one of the prefixed commands, and only then does it handle the escalated path. Ordinary requests to the host do not activate any of it, which is stated twice, once in the feature list and once in the workflow paragraph. The payoff for that restraint is the gate: with a stated threshold of three quarters agreement, a disagreement between models surfaces before the work reaches production instead of after it.

Five of the providers cost nothing extra, and one changed its terms

The cost paragraph is unusually concrete. It names five providers as free to add if you already have access, and explains why for four of them: the coding agent CLI, the Copilot seat, the Cursor command line and one more all run on a subscription you already pay for or on local authentication, and the local model runner is free because it runs on your own machine. A sixth provider is the exception with a date: the Qwen seat now requires an API key or a coding plan authentication, because its free authorisation tier ended on a specific day in April 2026. That is a small paragraph with a lot of information in it: it tells you which integrations are worth wiring up on day one, and it tells you that provider terms change independently of the plugin.

The council is a structured deliberation with a veto, not a chat

There is one command that runs several models against each other deliberately, and its parameters are what make it a method rather than a conversation. It takes a goal, with modes for advice, a decision, a plan, an implementation and a review, and a style, with options including a balanced one, an adversarial one, a red team one, an executive one and an implementation one. The examples in the file are written the way you would write a question to a colleague:

/octo:council --goal decision --style adversarial "Should this service stay monolithic?"

The number of personas is selectable, and the gates are structural: a quorum rule, a critical veto rule, a budget cap, and a handoff into a separate worktree for a plan that passed.

The roster is a default, and your existing pins still win

The current release notes describe a default roster rather than a lock, and the distinction is the point. The largest model leads architecture, planning, security reasoning and final judgement, an independent model is the implementation and review peer, a middle model is the standard seat, and one more is an opt in escalation for judgement. Then the sentence that protects an existing installation: model pins and provider configuration you already have still win. Overriding is done through environment variables such as OCTOPUS_OPUS_MODEL and OCTOPUS_CODEX_MODEL, and through a command that inspects or changes the roster, including a tier argument for adding a single bounded seat. So the project assumes you already route models by hand and gives you a default rather than a policy, which is the right default for a plugin whose users are model pickers.

Thirty one personas, fifty four commands, sixty three skills

The inventory is three numbers and one design sentence. Thirty one specialised personas are role specific agents, such as a security auditor or a backend architect. Fifty four commands are the slash commands you type. Sixty three skills are reusable workflow modules. The design sentence is that explicit workflows select the experts they need and ordinary host requests do not activate any of it, which is the same restraint as the provider list.

Five plugin directories ship inside one package

The packaging tells you who this is for. One package publishes five plugin directories, one each for the host, for the Codex CLI, for Cursor, for a fourth tool and for OpenCode, and it also ships a model context protocol manifest and its own protocol server. The manifest's main entry is a shell script, and the declared engine is a specific minimum runtime version. The publish check is a line of code that refuses to publish unless the version field looks like a real version and two specific files exist, which is a cheap guard against publishing an empty package. The build file is where the discipline shows: a sync target that regenerates three derived artefacts, a check mode of the same target that continuous integration enforces, and a changed-files gate that fails closed into the full local suite when a shared or generated file is touched.

Editorial conclusion

The plugin suits someone who wants a second opinion on architecture, security reasoning or a plan, and who is disciplined about when to escalate, because the escalation is explicit by design. It does not suit a team that wants every answer arbitrated by a panel, since the default path stays on one model and the extra providers cost time when they do answer. Before enabling the premium routes, read the routing strategy document, and check which providers your existing subscriptions already cover, since the readme names five that cost nothing extra.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes