Articles/Architecture Studio Now Runs on Codex
Architecture Studio Now Runs on Codex
Architecture Studio v1.4.2 brings all 46 open-source AEC skills to Codex, with native plugin packaging, portable workspaces, and the same project memory used in Claude Code.
Architecture Studio began as a public question: what does an AI workbench look like when architects can inspect it, change it, and own the judgment inside it?
Until this week, our answer ran in Claude. Architecture Studio v1.4.2 adds Codex as a supported host. The same 46 skills now install as one plugin in Codex or Claude Code, read the same project records, and write into the same user-owned studio workspace.
This is not a port to a second catalog. It is one catalog with two native entry points.
The numbers
As of release:
- 306 GitHub stars and 66 forks
- 46 shared skills
- 11 releases since the public launch
- One plugin identity:
as@skills-for-architects
The repo remains public, MIT-licensed, and made mostly of files an architect can open and read.
What compatibility means
Codex and Claude Code do not invoke skills the same way, store local extensions in the same directory, or expose the same automation. We did not flatten those differences behind documentation that only works on one side.
| Codex | Claude Code | |
|---|---|---|
| Studio entry point | $studio | /as:studio |
| Guided course | $learn | /as:learn |
| Skill menu | $tool-catalog | /as:tool-catalog |
| Firm skills | .agents/skills/ | .claude/skills/ |
| Workspace instructions | AGENTS.md | CLAUDE.md |
| Shared catalog | 46 skills | 46 skills |
| Host-specific automation | Explicit skill sequences | 7 native agents and 4 hook handlers |
Every bundled skill now carries a small host contract: use $skill-name in Codex and /as:skill-name in Claude Code. Supporting scripts and references resolve from the skill that was actually loaded, rather than assuming a Claude-only plugin path.
That sounds like plumbing because it is. Plumbing is what makes “compatible” mean the workflow runs, not just that a Markdown file can be copied into another directory.
The package is native to Codex
The repository now contains a Codex plugin manifest and a Git marketplace entry alongside the existing Claude package. That gives Codex a real installation surface for the whole catalog instead of asking users to symlink 46 skills by hand.
It follows the same underlying model described in OpenAI’s skill documentation: a skill is a readable package of instructions, references, assets, and optional scripts, loaded when the task calls for it. Architecture Studio supplies the AEC layer on top — source expectations, professional boundaries, project records, and maintained workflows for the work architects already do.
The workspace stays portable
Architecture Studio 1.4 moved the important state out of the plugin cache and into a workspace the firm owns. Codex compatibility preserves that boundary.
A new studio now includes both AGENTS.md and CLAUDE.md, plus .agents/skills/ and .claude/skills/. A project still keeps its facts in PROJECT.md, its reasoning in decisions/, and its meetings, site reports, plans, tasks, time, product data, and EPD data in plain files underneath it.
Those records do not become Codex records or Claude records. They remain project records. A firm can change hosts without migrating its memory into another product-shaped database.
Existing v1.4 workspaces remain valid. The new scaffold adds host-specific instruction and skill roots; it does not move or duplicate user-owned work.
What we did not pretend was portable
Claude Code’s native agents and hooks are Claude Code features. The Codex package does not register those files under new names and call that parity.
Instead, $studio has explicit Codex routes for the same seven workflow lanes: site context, NYC zoning, workplace programming, specifications, materials and FF&E, sustainability, and presentations. It invokes the relevant shared skills in sequence and requires the same synthesis and professional-review boundaries.
The output can be equivalent without claiming the orchestration mechanism is identical.
The guided course follows the same rule. $learn and /as:learn use one six-module fictional museum project, but the instructions branch where Codex uses AGENTS.md, .agents/skills/, and a new session while Claude Code uses CLAUDE.md, .claude/skills/, and /clear.
Install it
In Codex:
codex plugin marketplace add AlpacaLabsLLC/skills-for-architects
codex plugin add as@skills-for-architects
codexThen try:
$studio 123 Main St, Brooklyn NYOr, if this is your first time working this way:
$learnClaude Code continues to use the same plugin identity:
claude plugin marketplace add AlpacaLabsLLC/skills-for-architects
claude plugin install as@skills-for-architectsThe installation guide covers both hosts. The product page shows the complete system, and the source remains the place to inspect every workflow before trusting it with project work.
Why this matters
Architecture firms should not have to rewrite their practice layer every time they change the model or harness underneath it.
The valuable part is not the sigil before a command. It is the judgment encoded in the skill, the evidence attached to the output, the project memory that survives the session, and the firm’s ability to inspect and revise all three.
Codex compatibility is one more step toward making that layer portable. Not hostless — the hosts are meaningfully different — but owned above the host.