ArcKit: an enterprise architecture governance harness that runs inside your AI coding assistant

The governance paperwork problem ArcKit is aimed at

Enterprise architecture work produces a predictable pile of documents: principles, stakeholder maps, risk registers, business cases, requirements specifications, data models, design reviews, vendor evaluations and traceability matrices. Each one usually lives in a different template, a different folder and a different review cycle. The README frames ArcKit as a way to turn that scattered output into a systematic, AI-assisted workflow across strategy, design, delivery and assurance.

The audience is narrow and specific. The topic list points at enterprise architecture, procurement, public sector, govtech and compliance. The command set leans on UK Government frameworks by name: HM Treasury Orange Book for risk and HM Treasury Green Book SOBC for business case justification. Regional and sector overlays cover UAE, France, Canada, EU, Austria, Australia, US, UK-NHS and UK-GCloud. If you are an architect in a commercial SaaS company with no formal assurance process, most of this structure is overhead you will never invoice anyone for.

The README is also explicit about the boundary of the output: ArcKit produces draft artefacts for qualified people to review, not legal, regulatory, clinical or security advice. That sentence should shape how you read everything else. The tool generates the shape and the first pass of the document. A named human still owns the content.

How the plugin and overlay architecture actually works

ArcKit is not a single application. It is a set of plugins distributed through a Claude Code marketplace, plus a Python CLI package named arckit-cli, plus a JavaScript dependency set for diagram rendering. The README describes a core plugin called arckit containing 75 commands for UK Government civilian and generic enterprise work, and 14 community plugins layered on top.

The layering is compositional rather than flat. The README states that the 14 community plugins require the arckit core plugin, and that arckit-au-energy additionally requires arckit-au, which it composes. So installing the energy sector overlay without the Australian jurisdiction overlay is not a supported configuration. The README gives the exact install line for that combination.

There are two special cases worth reading twice. arckit-uk-gcloud is described as a proprietary, Claude Code only supplier-side G-Cloud bid-authoring overlay: public for installation and inspection, but not MIT licensed and not distributed to the non-Claude extension formats. arckit-fde is a tooling plugin with a single command, /arckit-fde:create, that generates a brandable Forward Deploy Engineering consulting website into docs/, ready for GitHub Pages, with UK Public Sector and Generic market presets. Neither behaves like a governance doc-type plugin.

Claude Code is named as the primary development platform and the only one that gets the full set: all official commands, autonomous research agents, automation hooks, bundled MCP servers (AWS Knowledge, Microsoft Learn, Google Developer Knowledge, govreposcrape, uk-tenders) and marketplace updates. The repository also carries an extensions/ directory and the README refers to non-Claude extension formats, so other assistants are supported, but the README does not enumerate what each non-Claude format loses.

Installing ArcKit in Claude Code and running a first command

The README states the Claude Code plugin path requires v2.1.251 or later, and it explains why in unusual detail: that release stops the file tools following a symlink swapped inside the working directory after the permission check, and makes Grep and Glob honour Read() deny rules through a symlinked search path. ArcKit's file-protection and secret-file-scanner gates sit in front of that class of bypass, so the floor is a security floor, not a cosmetic one.

First, update the assistant itself. The README gives this command:

claude install latest

Then add the plugin marketplace from inside Claude Code. The README shows this as a slash command:

/plugin marketplace add tractorjuice/arckit-claude

After that, install from the Discover tab, or use the CLI. The core plugin is the minimum viable install, and the README gives the exact package reference:

claude plugin install arckit@arckit-claude

If you need a jurisdiction, add it alongside the core plugin rather than instead of it. The README's example for the broad overlay set is:

claude plugin install arckit arckit-{uae,fr,ca,eu,at,au,us,uk-nhs,uk-gcloud}

The README states the marketplace ships 17 plugins and advises installing only the overlays you need. That advice is worth taking literally: arckit-uk-gcloud carries a different licence from the rest, so pulling the full set in one line means accepting a proprietary component you may not want.

The CLI is a separate package. According to pyproject.toml, arckit-cli requires Python 3.11 or later, declares its licence as MIT, and exposes a console script named arckit. The README does not document a pip install line for it, so check the project's own installation documentation before assuming the command name is enough to get it onto your machine.

Where ArcKit stops being the right tool

The most important limitation is stated by the project itself rather than discovered by a user: the output is a draft for qualified review. Nothing in the README suggests ArcKit validates a business case against Treasury guidance, checks a risk register for completeness, or signs off a design. It produces the artefact. If your organisation treats generated documents as the deliverable rather than the input to a review, ArcKit will accelerate the production of unreviewed paperwork, which is worse than producing less of it.

The second limitation is platform coupling. Claude Code is described as the primary development platform and the only one with the complete feature set. The README does not say what a Codex CLI, Gemini CLI, GitHub Copilot or opencode user gets, only that non-Claude extension formats exist. If your organisation has standardised on a different assistant, you are adopting an unknown subset.

The third is the version floor. Requiring a specific assistant build for security reasons means ArcKit inherits that assistant's release cadence. The README's own account of v2.1.246 fixing four plugin-loading bugs that hit ArcKit's exact layout shows how tightly the two are coupled: a plugin layout with skills under skills/*/SKILL.md broke on a specific assistant version. Pinning one side pins the other.

Finally, the bundled MCP servers include two keyed servers whose API keys sit in request headers, and the README notes that on a session with no keys configured those connections fail by design. Expect error output on a fresh install. It is not a broken setup.

ArcKit against a plain document template repository

The realistic alternative is not another AI governance product. It is a repository of Markdown templates plus a wiki, which is how most architecture practices already operate. The difference in approach is that a template repository gives you a blank structure and leaves the drafting, cross-referencing and consistency checking to the architect. ArcKit gives the assistant a command per artefact, so the drafting step is performed by the model against the command's instructions, and the architect's time moves to review.

That trade is not automatically favourable. Templates are deterministic: the same template produces the same structure every time, and a reviewer can diff two documents and see what changed. Generated artefacts vary with the prompt, the model and the context the assistant gathers. ArcKit's answer to that is its traceability features, including citation traceability with inline [DOC-CN] markers carrying source quotes, and requirements traceability across documents. Those are the mechanisms that make generated output auditable, and they are the parts worth evaluating before anything else.

A second alternative is a dedicated requirements management or architecture repository tool. Those enforce a schema and a workflow at the database level. ArcKit enforces structure at the prompt and file level, which is lighter to adopt and easier to abandon, but also easier to bypass. If your assurance process depends on the tool refusing an incomplete submission, ArcKit is the wrong shape.

Maintenance, licensing and the upgrade cost you are signing up for

The repository is not archived, and the last push was on 2026-09-03. The release history shows v6.12.0 on 2026-08-19, v6.13.0 on 2026-08-31 and v6.14.0 on 2026-09-03, so the cadence has been roughly weekly to fortnightly across that window. pyproject.toml sets version 6.14.0 and classifies the project as Development Status 5, Production/Stable, with Python 3.11 and 3.12 classifiers.

Upgrade cost is the part most teams underestimate. The README's version-floor discussion describes a chain of assistant releases, each fixing something ArcKit depends on: MCP diagnostics printing resolved secrets, WebSearch returning a 400 at high effort settings with thinking disabled, PreToolUse auto-allow hooks bypassing tool restrictions inside background agent tasks, and Bash permission-check bypasses. The README states that 18 ArcKit commands run at effort max. If your assistant build predates those fixes, those commands and the research agents are affected. Upgrading ArcKit therefore means tracking the assistant's release notes as well as ArcKit's.

On licensing, the picture is mixed and the README is explicit about only part of it. The README badge says MIT, pyproject.toml declares MIT for the CLI, and the repository LICENSE file is present. The GitHub API reports the licence as NOASSERTION, which means the platform could not classify it automatically. Separately, the README states that arckit-uk-gcloud is not MIT licensed and is not distributed to non-Claude extension formats. Treat the plugin set as having more than one licence and read the specific component before you install it. This is a description of what the files say, not legal advice.

Editorial conclusion

ArcKit fits enterprise architects, delivery leads and assurance teams who already work in a supported AI assistant and need a repeatable structure for principles, requirements, business cases and design reviews, particularly in UK public-sector or regulated contexts where the Green Book, Orange Book and traceability matter. It does not fit teams without a human reviewer for every draft, teams expecting a standalone web application, or anyone who needs a signed-off legal, clinical or security opinion from a tool. Before adopting it, confirm your assistant meets the stated Claude Code v2.1.251+ floor, decide which of the 17 plugins you actually need, and check the licence file for the components you install, since the README states that arckit-uk-gcloud is proprietary and not MIT licensed.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. Releases
5. tractorjuice/arc-kit on GitHub