ArcKit: A Governance Harness That Runs Inside Your AI Coding Assistant
The Enterprise Architecture Governance Harness — strategy, architecture, delivery, and assurance using AI coding assistants
At a glance
- What is it?
- ArcKit is a set of Claude Code plugins that turn enterprise architecture artefacts (principles, business cases, requirements, design reviews, traceability) into slash commands and agent workflows. It is opinionated toward UK public sector templates, and the README is explicit that its output is draft material for human review.
- Who is it for?
- ArcKit suits architects already working inside Claude Code who need repeatable, template-driven artefacts: principles, Green Book SOBC business cases, Orange Book risk registers, HLD/DLD review packs, and requirement traceability. It is a poor fit if your team standardises on Codex CLI, Gemini CLI, or Copilot as the primary assistant, or if you need a governance tool with its own UI, database, and audit log rather than files in a repository.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 12 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What ArcKit Is Actually Shipping
ArcKit is not an application you deploy. It is a collection of prompt assets, commands, and agent definitions that load into an AI coding assistant and produce architecture documents as files in your working directory. The README describes it as a harness for strategy, design, delivery, and assurance workflows, and lists the artefacts it can generate: architecture principles, stakeholder driver analysis, HM Treasury Orange Book risk material, HM Treasury Green Book SOBC business cases, requirements documents, data models with ERD and GDPR considerations, Wardley Maps, Mermaid diagrams, vendor RFP and selection material, and formal HLD/DLD design reviews. The primary language in the repository metadata is JavaScript, which is consistent with plugin and tooling glue rather than a compiled runtime. The intended user is an enterprise architect, particularly one in or adjacent to UK government or a regulated sector, who already uses an AI assistant for drafting and wants the drafting to follow a named method instead of starting from a blank chat.
The Plugin and Overlay Model
The architecture is compositional. A core plugin named arckit carries what the README calls 75 commands covering UK Government civilian and generic enterprise work. Everything else is an overlay that requires the core: regional packs for UAE, France, Canada, EU, Austria, Australia, and the US, sector packs such as UK NHS, UK finance, and Australian energy, plus arckit-togaf-adm for TOGAF ADM and arckit-agent-architecture for AI agent governance. The README states that arckit-au-energy composes arckit-au, so an Australian energy install is three plugins. This is a real design decision with a real cost: the more overlays you install, the more command surface your assistant carries, and the README does not describe a mechanism for resolving conflicting guidance between an overlay and the core. The distribution channel is a Claude Code marketplace hosted at tractorjuice/arckit-claude, and the README notes the older tractorjuice/arc-kit marketplace remains for compatibility while new installs should use the newer one.
Installation and the Version Floor
The documented install path assumes Claude Code. The README gives the sequence: run claude install latest, add the marketplace with /plugin marketplace add tractorjuice/arckit-claude, then install from the Discover tab or the CLI. A core install is claude plugin install arckit@arckit-claude. Overlays are added by name, for example claude plugin install arckit arckit-togaf-adm arckit-agent-architecture. The README states a minimum of Claude Code v2.1.251, and the reasoning is unusually specific: earlier releases allowed file tools to follow a symlink swapped after the permission check, and ArcKit ships file-protection and secret-file-scanner gates that sit in front of exactly that class of bypass. The same release also downgrades Opus 5 effort settings so that ArcKit's 18 effort: max commands complete when thinking is off. The README further notes that v2.1.246 fixed plugin-loading bugs affecting ArcKit's layout, including skill discovery under skills/*/SKILL.md and duplicate SHA-named cache directories. Treat the floor as a hard requirement, not a suggestion: on an older client, the README's own account is that commands may fail or gates may not behave as documented.
MCP Servers and What Fails by Design
ArcKit bundles MCP servers, which is where its research commands get their sources. The README names AWS Knowledge, Microsoft Learn, Google Developer Knowledge, govreposcrape, and uk-tenders, and describes Azure research as using Microsoft Learn MCP for authoritative documentation. Two of those servers are keyed, meaning their API keys are supplied through user_config and travel in request headers. The README is candid that on a session with no keys configured, those connections fail by design. That is a meaningful operational detail: a fresh install without credentials will show failed MCP connections, and the README ties its version floor partly to a Claude Code release that stopped MCP diagnostics printing resolved secrets. If you run ArcKit in a shared or logged environment, that history is the reason to keep the client current rather than pinning an older build for stability.
Citation Traceability and the Draft-Only Boundary
Two features distinguish ArcKit from generic document prompting. The first is citation traceability: the README describes inline [DOC-CN] markers paired with source quotes for external documents, so a generated claim can be traced back to a passage. The second is requirements traceability as a first-class concern, listed alongside the requirements documents themselves. Both are the kind of thing that usually lives in a requirements management tool, and here it lives in generated text. The README also draws an explicit boundary: ArcKit produces draft artefacts for qualified people to review, not legal, regulatory, clinical, or security advice, and it links to a section titled What ArcKit does not do. Take that at face value. A Green Book SOBC or an Orange Book risk register generated by an assistant is a starting structure, and the review step is where the governance value actually sits. If your process treats generated artefacts as approved outputs, ArcKit's own documentation argues against you.
Where the Model Breaks Down
The clearest limitation is platform coupling. The README calls Claude Code the primary development platform and states that the proprietary arckit-uk-gcloud overlay is Claude Code only and not distributed to the non-Claude extension formats. The topics list mentions Codex CLI, Gemini CLI, GitHub Copilot, and OpenCode, so some form of multi-assistant support exists, but the README does not describe feature parity, and the richest capabilities (autonomous research agents, automation hooks, bundled MCP servers, automatic updates via the marketplace) are described as Claude Code features. A second limitation is licence ambiguity. The repository metadata reports NOASSERTION while the README badge and text describe MIT for the core, and the README states plainly that arckit-uk-gcloud is public for installation and inspection but not MIT licensed. If you are assembling a dependency inventory, that mix needs a human decision, not a badge. A third is that overlays encode jurisdiction-specific templates. Installing arckit-us does not make the core's UK Treasury references disappear, and the README offers no guidance on which combination is coherent for a given organisation.
How This Differs From a Requirements Management Tool
The obvious alternative is a conventional requirements and architecture repository such as a dedicated ALM or EA platform, where artefacts live in a database with roles, workflow states, and an audit trail. The difference in approach is not cosmetic. ArcKit keeps artefacts as files in your repository, generated on demand by an assistant, versioned by whatever you already use for the repository. That gives you diffable output and no new system to operate. It also gives you no enforced workflow, no permission model, and no queryable traceability graph beyond what the generated text and markers provide. If your governance depends on proving who approved which version of a principle on which date, a file-based harness does not answer that question by itself. If your governance depends on producing a well-structured first draft quickly and then reviewing it, the file-based approach is the lighter one.
Maintenance Cost and Release Cadence
The release history shows v6.12.0, v6.13.0, and v6.14.0 within roughly two weeks in August and September 2026, with the last push on the main branch on 2026-09-03. That cadence is the maintenance story: the project tracks Claude Code releases closely, and the README's version-floor discussion reads like a changelog of client bugs that affected ArcKit directly. For an adopter, that means upgrade work is ongoing rather than occasional. The marketplace provides automatic updates, which lowers the mechanical cost but raises the question of when a governance-sensitive workflow changes underneath you. The README does not describe a deprecation policy for commands or overlays, so pinning is a choice you would have to make outside the documented flow. On licensing, the core is described as MIT while the repository metadata says NOASSERTION and one overlay is proprietary; confirm the actual terms for each plugin you install rather than assuming the badge covers the set.
Editorial conclusion
ArcKit suits architects already working inside Claude Code who need repeatable, template-driven artefacts: principles, Green Book SOBC business cases, Orange Book risk registers, HLD/DLD review packs, and requirement traceability. It is a poor fit if your team standardises on Codex CLI, Gemini CLI, or Copilot as the primary assistant, or if you need a governance tool with its own UI, database, and audit log rather than files in a repository. Before adopting, verify three things yourself: that your Claude Code version is at or above v2.1.251, that every overlay you install is actually MIT-licensed (the README states arckit-uk-gcloud is proprietary and Claude Code only, and the repository licence field reads NOASSERTION), and that your organisation accepts AI-generated drafts as inputs to a review process that humans still own.
Community notes