Arch Studio 1.5: The Four Layers a Firm Builds On

Arch Studio 1.5 is out on both channels. The same release runs as the open-source plugin in Codex and Claude Code, and as the hosted MCP inside your Claude or ChatGPT desktop or mobile app. One version, two ways in.

The numbers

The Codex post went out on August 12 at 306 stars. Since then:

• 358 GitHub stars and 72 forks
• 61 shared skills, up from 46, still MIT-licensed, still free
• Fifteen releases, four of them since that post: 1.4.3, 1.4.4, 1.4.5 and now 1.5.0
• About 1,400 clones by 810 people in the last two weeks; the repository page had 1,500 unique visitors in the same window

We had real help from the community: four issues and a pull request from people we had never met. Two of the four releases above, 1.4.4 and 1.4.5, trace back to one of those reports.

Five ways this release changes Arch Studio

1. A simplified architecture in four layers
2. Practice areas
3. A hosted MCP, so a firm can use it without installing anything
4. A clearer contract between Arch Studio and the harness: your assistant does the work, with its own tools
5. Norma, a coordinator that plans the work and hands it to the right skill or specialist

And fifteen new skills out of the box: product URL capture, spec book assembly, quantity take-off from drawing sets, product audits and cut sheets, proposals, agreements and invoices, a document register, and more.

A simplified architecture

Governance. A framework for AI studio governance that combines agent rules, skill-building patterns, a simple contract between Arch Studio and the harness, the definition of your studio, and the hooks, access, checks and task-completion rules. Studio rules · Agent rules · Patterns

Knowledge. A framework for linking to the external knowledge critical to the everyday practice of architecture, from codes to standards, best practices, industry schemas and more, scalable across geographies and practice areas. Arch Studio carries the links and the routing, never a copy or a summary; the original is what gets read. Codes · Standards · Reference datasets

Execution. A framework for defining what is a skill, a tool or an agent and how they work together to get a task done. In 1.5 every skill is a complete specification of outcomes, schemas, checks and evidence, and your assistant performs it with the tools it already has. No runner, no installer, no bundled script. Skills & tools · Agents

Memory. A framework for keeping project records, decisions, even that website you once asked Arch Studio to fetch a product from, automatically and locally, all with provenance so you know where it came from. Records are plain CSV registers in your own folders: documents, tasks, time, invoices, changes. Project records · Decisions · Product captures

Practice areas

Architectural knowledge already organizes itself this way, and it helps us define a specific part of the craft. When tools, references and experience in the form of memory are combined, a studio can develop a particular area of the craft it is especially good at. Some will build practice areas around sustainability or multifamily development, others around code research or fabrication. Whatever the case, capturing that knowledge needed an organizational structure that was missing from Arch Studio until now.

Arch Studio MCP

Why an MCP? The Claude Code or Codex plugin is a good delivery method for this kind of project, but it takes some technical knowledge to install and stand up. MCPs have quickly become the de facto way for apps to connect with Claude or ChatGPT, and through them we reach a much wider audience. By doubling down on Claude and ChatGPT as the main surface Arch Studio lives in, we also get their continued development of critical capabilities, and Arch Studio sits within reach of the other connections you already have there: email, calendar, file storage, design tools like Figma and Miro, Airtable, Notion and many more. The hosted MCP delivers the same 1.5 release as the plugin, every one of the 61 skills, and is free for firms. Pick one or the other; you do not need both.

A clearer contract between Arch Studio and the harness

Your desktop apps for Claude or ChatGPT are what the AI industry calls harnesses. They serve you your favorite LLM, but with a large amount of tooling that makes that LLM far more capable. These apps are built for the general public and their capabilities improve constantly. Only a few months ago your local app could not browse the web very effectively. This summer both shipped a built-in browser: ChatGPT on July 9, when OpenAI folded Atlas and Codex into one desktop app, and Claude on August 26, inside Cowork in the desktop app. The agent can now open any website you point it to, read it, click through it and type into it.

That is why 1.5 stops shipping code for the assistant to run. Earlier releases bundled scripts for the mechanical steps; they worked in some hosts and not in others, and every host that could not run them got a worse Arch Studio. Now each skill states what a correct result looks like and how to check it, and the harness does the work with whatever it has. The trade is visible: a workflow that used to call a script takes more turns to do the same job, because the assistant now inspects, prepares, verifies and applies instead of calling one command. What we get back is an Arch Studio that improves every time Anthropic or OpenAI ship, as well as every time we do.

Norma

Norma is the coordinator. Describe the work in plain language and Norma plans it, hands each part to the right skill or specialist agent, and keeps the record straight. In the plugin, Norma sits alongside the seven specialists as the eighth agent; through the MCP, Norma is how you talk to Arch Studio at all. Asking for a specific skill directly still works exactly as before.

The process

Six weeks and a few billion tokens went into this release, and a lot of learnings from Sharmila Chaudhuri’s Studio, a small architecture design studio, and Nicolás Martínez D’Alto’s team at Slantis, a mid-sized architecture production studio. Working with each helped us shape the architecture firms build their own AI studio practice on. Many of the FF&E skills in this release, from URL capture to the spec book, come straight from their projects.

Thank you, Sharmila and Nico, for the projects and the notes. What comes next only a firm can do: your standards, your procedures, your decisions, on top of the four layers.

Sign in at architecturestudio.ai if you run a firm. Fork it if you want to make it yours. Read more on the product page or in the install guide. Upgrading from 1.4: new studios and projects use CSV registers, and existing workspaces are not converted, so keep a 1.4 workspace on 1.4.5 until you set it up again with 1.5.