Multi-platform AI agent orchestration for GitHub Copilot, Claude Code, and Codex. It installs specialist agents, shared skills, memory templates, hooks, and project instructions.

Install

npx agent-arche install

Run it from your project root. The installer asks for:

Scope: Full orchestration, Small orchestration, or Skills only
Platform: GitHub Copilot, Claude Code, or Codex
Copilot plan, only when installing Copilot agents

Scope	Installs
Full orchestration	Full specialist roster, hooks, instructions/rules, prompts/commands, skills, memory, and platform root files
Small orchestration	Lean 3-agent setup: Orchestrator, Coder, Docs-updater, plus hooks, instructions/rules, skills, memory, and root files
Skills only	Shared `skills/` folder only

Platform	Destination
GitHub Copilot	`.github/`
Claude Code	`.claude/` plus root `CLAUDE.md`
Codex	`.codex/`, `.agents/skills/`, plus root `AGENTS.md`

After installing Full or Small orchestration, run the project startup flow once:

Platform	Prompt
GitHub Copilot	`@Orchestrator use the project-startup skill on this project`
Claude Code	`Use the orchestrator agent to run the project-startup skill on this project`
Codex	`Use the orchestrator agent to run the project-startup skill on this project`

Update later with:

npx agent-arche update

Orchestration

Full orchestration uses specialist agents. Small orchestration keeps the same skills and memory model but routes most work through Coder and Docs-updater.

Agent	Full orchestration purpose	Small orchestration
Orchestrator	Classifies work, confirms the agent pipeline, delegates, and never edits files directly	Same coordinator, lower-token routing
Planner	Researches the repo and creates implementation plans	Folded into Coder or Orchestrator
Researcher	Investigates prior art, docs, issues, CVEs, and unknown libraries	Folded into Coder
Coder	Implements code and unit tests	Main executor
Designer	Handles UI/UX implementation and visual direction	Folded into Coder with `design`
Code-reviewer	Reviews code quality, conventions, TypeScript, and maintainability	Folded into Coder or direct skill use
Security-auditor	Reviews auth, routes, secrets, injection, SSRF, file I/O, and OWASP risks	Folded into Coder or direct skill use
UX-reviewer	Reviews accessibility, usability, interaction, and visual quality	Folded into Coder with `design`
Tester	Writes and runs Playwright flows	Folded into Coder
Docs-updater	Updates memory, docs, changelogs, commit text, and PR text	Same docs and memory owner

Common full pipelines:

Work	Pipeline
New feature	Planner -> Researcher -> Coder + Designer -> Code-reviewer + Security-auditor + UX-reviewer -> Tester -> Docs-updater
Bug fix	Planner -> Coder -> Code-reviewer -> Tester -> Docs-updater
UI change	Designer -> Code-reviewer + UX-reviewer -> Docs-updater
Architecture review	Planner with `improve-codebase-architecture` -> Docs-updater

Skill Routing

The public skill surface is the set of top-level folders under skills/ that contain SKILL.md. Design subflows live inside skills/design/, and caveman compression lives inside skills/caveman/.

Work type	Full owner	Small owner	Skills
Bootstrap and memory	Orchestrator, Planner	Orchestrator	`project-startup`, `caveman`
Product planning	Planner	Orchestrator or Coder	`product-brainstorming`, `create-prd`, `create-epics-and-stories`
Planning pressure tests	Planner, Coder	Orchestrator or Coder	`grill-me`, `grill-with-docs`, `triage`
Implementation and debugging	Coder, Tester	Coder	`diagnose`, `tdd`, `prototype`, `handoff`
Architecture and review	Planner, Code-reviewer	Coder	`improve-codebase-architecture`, `review`
Code standards	Coder, Code-reviewer	Coder	`coding-standards`, `karpathy-guidelines`, `api-design`, `postgres-patterns`, `seo`
UI/UX	Designer, UX-reviewer	Coder	`design`
Git and PRs	Coder, Code-reviewer, Docs-updater	Coder or Docs-updater	`git`

Recommended order for product work: project-startup once, product-brainstorming, create-prd, then create-epics-and-stories.

Skills

Skill	Purpose	Example request
`api-design`	Shapes REST APIs, validation, pagination, error envelopes, status codes, URL design, and API security checks.	"Design the endpoint shape and error responses for this checkout API."
`caveman`	Compressed communication mode and context-file compression workflow via `steps/compress.md`.	"Use compressed mode" or "compress AGENTS.md with caveman."
`coding-standards`	General code quality rules for TypeScript, naming, file organization, imports, errors, and forbidden patterns.	"Review this module against our coding standards."
`create-epics-and-stories`	Turns a PRD into low-dependency epics, stories, acceptance criteria, and build slices.	"Break this PRD into epics and stories for the dev team."
`create-prd`	Creates a PRD through a guided workflow and reusable template.	"Create the PRD for this product idea."
`design`	Consolidated UI/UX skill for product-fit direction, implementation, redesign, animation, GSAP, styles, audits, critique, optimization, Stitch, and full-output flows.	"Build a polished settings page and audit the checkout flow."
`diagnose`	Debug loop for reproduce, isolate, hypothesize, instrument, fix, and regression-test work.	"Diagnose this flaky login test with evidence."
`git`	Branch naming, commit conventions, PR standards, and PR body guidance.	"Create the branch name, commits, and PR body for this change."
`grill-me`	Pressure-tests a greenfield idea before code or docs exist.	"Interrogate this idea before I start building."
`grill-with-docs`	Pressure-tests plans against existing code and docs, then records decisions as context or ADRs.	"Stress-test this plan against the current architecture docs."
`handoff`	Produces compact continuation context for another agent, person, or session.	"Create a handoff for the next agent."
`improve-codebase-architecture`	Finds shallow modules and architecture friction, then proposes deeper module boundaries and reports.	"Find shallow modules in this subsystem and propose refactors."
`karpathy-guidelines`	Coding-agent behavior guardrails: keep changes surgical, avoid overengineering, expose assumptions, and verify results.	"Keep this refactor surgical and call out assumptions before editing."
`postgres-patterns`	PostgreSQL schema, query, migration, indexing, RLS, pooling, and review patterns.	"Review this migration for indexes, RLS, and query performance."
`product-brainstorming`	Structured product ideation using guided creative techniques and output organization.	"Help me brainstorm the first version of this product."
`project-startup`	First-run setup that bootstraps memory and configures `docs/agents/` issue tracker, labels, and domain-doc conventions.	"Run the project startup flow for this repo."
`prototype`	Builds disposable logic or UI experiments to answer a product or technical question quickly.	"Prototype this workflow before we commit to the architecture."
`review`	Reviews a branch, PR, or diff against standards, specs, risks, and missing tests.	"Review this PR using the engineering review skill."
`seo`	SEO execution plan for technical foundation, metadata, structured data, Core Web Vitals, linking, and monitoring.	"Prepare this marketing page for search and sharing."
`tdd`	Guides red-green-refactor work with support for deep modules, interfaces, mocks, and refactoring candidates.	"Implement this parser with TDD."
`triage`	Moves issues through canonical states and can produce ready-for-agent briefs.	"Move issue 42 to ready-for-agent with an agent brief."

Runtime Pieces

Piece	What it does
Hooks	`session-start.cjs` injects startup context, `pre-tool-safety.cjs` blocks destructive commands, and `changelog-reminder.cjs` reminds agents about docs/changelog drift.
Instructions/rules	Copilot uses native instruction files. Claude and Codex install equivalent rule/instruction files for TypeScript, Svelte, tests, and API routes.
Memory	Full and Small orchestration install a project memory folder. Use `project-startup` once to seed memory and engineering skill context. Skills-only installs do not include memory.
Prompts/commands	Copilot prompts and Claude commands include `code-review`, `write-tests`, `debug`, `create-issue`, `pr-description`, and `security-review`. Codex uses agents by name instead.
Codex MCP	Codex MCP servers and runtime behavior are configured in `.codex/config.toml`, with disabled examples for Context7, Playwright, GitHub, and OpenAI docs.

Manual Setup

The CLI is preferred. For manual installs:

Platform	Copy
Copilot	`copilot/` to `.github/`, `skills/` to `.github/skills/`, `memory/` to `.github/memory/`
Claude Code	`claude/` to `.claude/`, `skills/` to `.claude/skills/`, `memory/` to `.claude/memory/`, `claude/CLAUDE.md` to root `CLAUDE.md`
Codex	`codex/` contents to `.codex/`, `skills/` to `.agents/skills/`, `memory/` to `.codex/memory/`, `codex/AGENTS.md` to root `AGENTS.md`

For PR and issue workflows, copy templates/ into .github/ and update the comments inside.

Notes

Copilot installs adjust agent models by subscription: Student replaces Claude models with Copilot GPT, Pro replaces Opus-only agents with Sonnet, and Pro+ keeps the configured roster.
Verify published package integrity with npm view agent-arche dist.integrity and compare it to agent-arche.json in the installed platform folder.

Credits

Built on top of excellent open-source work:

Repo	What we took from it
cyxzdev/Uncodixfy	Design and UI skill inspiration
pbakaus/impeccable	Foundation for the `ui-audit`, `ui-optimize`, `critique`, `animate` steps inside the design skill
Leonxlnx/taste-skill	`gsap`, `redesign`, `soft`, `minimalist`, `brutalist`, `stitch`, `output`, and other quality-focused steps inside the design skill
anthropics/claude-code frontend-design	Frontend design skill patterns
mattpocock/skills	`grill-me`, `grill-with-docs`, `tdd`, `improve-codebase-architecture`, `diagnose`, `handoff`, `prototype`, `review`, `triage`, and `setup-matt-pocock-skills` that is used inside the project startup skills
JuliusBrussee/caveman	`caveman`, and `caveman-compress` communication skills
BMad	`bmad-brainstorming`, `bmad-create-prd`, and `bmad-create-epics-and-stories` skills
Akindu23/my-agent-skills	`karapathy-guidelines`, `postgres-patterns` skills