
Octopus
The problem
Every team that uses an AI coding agent at scale ends up writing the
same kinds of configuration over and over: the project's coding
conventions, the team's commit conventions, the security guardrails,
the on-call playbooks, the post-mortem rituals. That config grows
tribal: a long CLAUDE.md, a folder of
.cursor/rules/*, a half-finished
copilot-instructions.md. Different agents speak
different dialects of the same idea.
Octopus is the configuration layer that sits above all of them. One source of truth β bundles, skills, hooks, roles, knowledge β that compiles down into the formats each agent expects. Add a new agent, change a coding rule, ship a new audit; one edit, all surfaces updated.
The five primitives
- Bundles β opinionated
packages of skills + roles + rules tuned to a team archetype
(
starter,quality,docs,backend,growth). - Skills β protocols
the agent engages by description match or explicit invocation
(
debug,doc-prd,refactor-deepen,audit-config, plus 30+ more). - Commands
β slash-command entry points that pin a specific skill to a typed
invocation (
/octopus:debug,/octopus:context-handoff, β¦). - Hooks β lifecycle scripts that enforce discipline automatically: format on save, block destructive git, propose CLAUDE.md updates at session end.
- Roles β personas
the user can delegate to (
@architect,@backend-developer,@writer, β¦) that scope an agent's behaviour for one task.
Install
macOS / Linux
curl -fsSL https://github.com/leocosta/octopus/releases/latest/download/install.sh | bashWindows (PowerShell)
irm https://github.com/leocosta/octopus/releases/latest/download/install.ps1 | iex
Then octopus update inside any repo to generate the
agent config files. Full walkthrough on the
Quick Start page.
Featured workflows
- TDD a new feature β vertical tracer-bullet slices, integration-style tests, hard ban on horizontal slicing.
- Pre-merge audit on a billing change β composes security + money + tenant + cross-stack contract audits in parallel.
- Hand off a session β compact the conversation into a tmp doc with prescriptive suggested-next-skills for the successor agent.