Project status — concluded experiment (July 2026). A3IP explored an AI-mediated install protocol for agent workflows. The design works, but the space it aimed at is now covered more maturely by Microsoft APM, an actively maintained dependency manager and installer for agent context. For real use, prefer APM. This site and its repositories remain as an honest, open record; the project is not being developed further.
A3IP logo — a suitcase with a travel stamp

A3IP — AI Infrastructure Installation Package · pronounced “ay-trip”

Package an AI workflow once — install it anywhere.

A permission-aware package format for portable AI agent workflows. Hand one bundle to your AI assistant and it installs the whole thing on Claude Code, Codex, Cursor, or Cowork — asking your permission first.

Browse packages Read the spec
Spec v1.12 Open standard · CC BY 4.0 pip install a3ip GitHub

What is this?

A standard for packaging AI workflows

You build a great AI workflow on one platform. You send it to a teammate on another — and it breaks: wrong paths, missing MCP setup, config keys with no context, no idea what to run first. Until now there has been no standard way to package an AI workflow so that another AI can install it from scratch, ask the right setup questions, and confirm the permissions it needs before touching anything.

Package format

Bundle skills, protocols, scripts, and artifacts into one .a3ip.bundle — shareable and installable across platforms.

Permission contract

Every filesystem path, network domain, MCP server, and shell command is declared upfront. A3IP doesn't sandbox — your AI is trusted to show the plan and honor it before acting. The value is that the contract is auditable.

Runtime-agnostic

The same bundle aims to install consistently across Cowork, Claude Code, Codex, and Cursor — the outcome depends on the receiving AI following the protocol. Per-platform mechanics live in adapters inside the package.

How is it different?

It composes with what you already use

A3IP adopts SKILL.md and composes with MCP and Cowork Plugins. It overlaps heavily with Microsoft APM, which already does portability and installation for developers — A3IP's narrower angle is a conversational, AI-run install. A3IP is not a runtime: your AI still runs the workflow; A3IP is the package format and the install contract.

vs. SKILL.md

A3IP packages SKILL.md as a native component — every A3IP skill already is a SKILL.md. A3IP wraps the install protocol around it.

vs. MCP

MCP is how an AI calls tools at runtime. A3IP sits above it: it declares which MCP servers a workflow needs and sets them up during a permissioned install.

vs. Cowork Plugins

Cowork Plugins are Anthropic’s first-party packaging for Cowork. A3IP is the cross-platform sibling — same discipline, no lock-in — and targets Cowork first-class.

vs. Microsoft APM

APM is a mature dependency manager — it resolves, pins, scans, and policy-checks agent context across seven harnesses, and installs it. A3IP's narrower angle is a conversational install the AI runs itself. The overlap is large.

Show me

One bundle. Installed by your AI.

No special support required — any capable AI agent that can read files and follow instructions can install an A3IP package. The bundle describes everything itself.

1
Hand your AI a .a3ip.bundleyouA single file — no shell scripts, no manual setup.
2
It reads INSTALL.md & CONFIGURE.mdaiLearns the steps, then asks you the setup questions it needs.
3
It presents the plan + permissionsaiEvery step and every permission declared upfront for you to review before it acts.
4
You confirm → workflow installedyouNothing runs until you say yes. Then it’s set up for your platform.

The permission contract your AI shows you — and waits on — before running a thing:

permissions:
  network:
    - domain: api.github.com
      reason: Read pull requests, post review comments
  filesystem:
    - path: ./reviews/
      access: write
      reason: Stores generated review reports
  shell:
    - command: python3
      reason: Runs the review script

The reference package gallery

Five packages you can install today — each a different shape of workflow.

a3ip-creator

Author a new A3IP package from scratch, guided end to end.

CoworkCodexClaude CodeCursor

ai-code-review-flow

Structured GitLab/GitHub code review on a 7-category checklist.

CoworkCodexClaude CodeCursor

ai-research-workspace

Capture papers, log experiments, synthesize findings — all local markdown.

CoworkCodexClaude Code

ai-standup-assistant

A daily standup compiled from your GitHub activity across repos.

CoworkCodexClaude Code

cowork-backlog

A lightweight backlog and read-only Kanban board across projects.

CoworkClaude Code
See the full gallery →

How do I start?

Install one, or build your own

Install a package

Grab a bundle URL from the gallery and paste it into a conversation with your AI. It reads INSTALL.md, walks you through CONFIGURE.md, and confirms every step before executing.

"Install this A3IP package:
 https://.../ai-code-review-flow-v1.5.0.a3ip.bundle"

Create a package

Scaffold a valid package in seconds with the CLI, validate it against the spec, and produce a distributable bundle.

pip install a3ip

a3ip scaffold my-workflow
a3ip validate my-workflow/
a3ip bundle   my-workflow/
✓ Cowork ✓ Codex ✓ Claude Code Cursor · community