Spin up a new software project optimized for three coding agents in parallel: Claude Code (Anthropic), Codex (OpenAI), Pi (pi.dev). One source of truth, three CLI overlays, one wiki, one verification gate.
# Scaffold a new project in any directory:
npx github:cubixhub/agent-framework my-new-appThat's it. The command will prompt you interactively for project name, branch,
PM tool (Linear / Plane / none), which CLIs to enable, and languages. It then
copies template/ to your chosen target, substitutes {{PLACEHOLDERS}}, runs
git init, makes an initial commit, and prints next-step instructions.
# Without specifying target (you'll be asked):
npx github:cubixhub/agent-framework
# Clone the repo first if you prefer:
git clone https://github.com/cubixhub/agent-framework.git
cd agent-framework
bash scripts/install-prereqs.sh # install claude, codex, pi, jq, tmux
bash scripts/init-project.sh ~/path/to/my-new-appinit-project.sh is interactive: it prompts for project name, default branch,
PM tool (Linear / Plane / none), enabled CLIs, and language stack. It then
copies template/ to your chosen target, substitutes {{PLACEHOLDERS}},
runs git init, makes an initial commit, and prints next-step instructions.
- Tri-CLI symmetry.
AGENTS.mdis the single source of truth; all three CLIs read it.CLAUDE.md,CODEX.md,PI.mdare thin overlays. Per-CLI config goes under.claude/,.codex/,.pi/— but agent runtime rules, skills, hooks, and memories all live under.agent-os/(shared). - Karpathy LLM-wiki. Every project gets a living
wiki/(sources → entities → concepts → questions → ideas → plan). Agents persist findings to the wiki on every session. No more re-deriving from raw on every query. - Symphony-style orchestration. A daemon under
orchestration/lets you run multiple agents in parallel against the same wiki + PM backlog, coordinating via wiki + Linear/Plane state, not by polling each other. - Verification-first. The 7-layer + 4-tier + 4-anti-gaming verification
stack (see
wiki/plan/verification-strategy.md) is scaffolded from day one. No work is "done" until verification returns PASS.
agent-framework/
├── README.md, INSTALL.md, CHANGELOG.md, LICENSE, DELIVERY.md
├── PROJECT-TEMPLATE-SPEC.md full design rationale
├── package.json + bin/agent-framework.js (npx entry point)
├── template/ copied verbatim into new projects
│ ├── AGENTS.md.tmpl universal entry doc (all 3 CLIs)
│ ├── CLAUDE.md.tmpl Claude Code overlay
│ ├── CODEX.md.tmpl Codex overlay
│ ├── PI.md.tmpl Pi overlay
│ ├── SYSTEM.md.tmpl Pi system prompt
│ ├── LICENSE.tmpl MIT (substitutes owner + date)
│ ├── .agent-os/ SHARED runtime config (rules, hooks, skills, memories)
│ ├── .claude/ Claude Code-specific overlay
│ ├── .codex/ Codex-specific overlay
│ ├── .pi/ Pi-specific overlay
│ └── wiki/ Karpathy LLM-wiki scaffold (SCHEMA, PLAN, IDEAS, STATUS, plan/)
├── scripts/ init-project.sh, install-prereqs.sh, verify-install.sh, skill-pull.sh
├── skills/ 17 skill groups, 57 files (skill-discovery, verification-first,
│ matt-pocock, ai-engineering, deep-research, factory-8, ...)
├── agents/ 22 role definitions (architect, implementer, scrutinizer,
│ parent-ai, wiki-curator, ml-engineer, ...)
├── orchestration/ Symphony-style poll-claim-process daemon
│ (tri-CLI adapters + Linear/Plane PM adapters)
├── integrations/ Linear + Plane bootstrap (workflow states + role labels)
├── docs/ 13 power-user guides rewritten for tri-CLI
└── examples/ quickstart-demo/ — real generated example project
- Power-user docs: see
docs/ - Skills library: see
skills/ - Agent roles: see
agents/ - Orchestration daemon: see
orchestration/ - PM adapters (Linear / Plane): see
integrations/ - Design rationale: see
PROJECT-TEMPLATE-SPEC.md - Completion summary: see
DELIVERY.md
MIT. See LICENSE.
Issues + PRs welcome at https://github.com/cubixhub/agent-framework