What is Archcore?

Shared

Archcore is a git-native context layer for AI coding agents. It stores project knowledge, such as specs, architecture, decisions, rules, and plans, as typed markdown documents in your repository. Agents read those documents from the first turn of a session.

Its two discovery categories are spec-driven development and context engineering. A spec is one part of context, not the whole context: the same directory also holds the architecture, prior decisions, constraints, and team rules an agent needs while implementing.

The problem

AI coding agents start every session from scratch. They do not remember yesterday’s decisions, last week’s architecture review, or the coding standards your team agreed on months ago.

Teams work around this with flat instruction files:

• CLAUDE.md grows into a wall of text that is hard to maintain.
• .cursorrules is locked to one tool and is not reusable across agents.
• A docs/ folder holds unstructured markdown, so agents cannot tell what is relevant.
• Team knowledge stays in people’s heads, Slack threads, and wikis.

These workarounds stop scaling as projects grow.

How Archcore solves this

Archcore adds a .archcore/ directory to your repository. Inside it, every piece of project knowledge becomes a typed document: an architecture decision, a coding rule, a how-to guide, or an implementation plan.

your-project/
├── .archcore/
│   ├── auth/
│   │   ├── jwt-strategy.adr.md        ← decision
│   │   └── auth-redesign.prd.md       ← requirements
│   ├── payments/
│   │   └── stripe-integration.guide.md ← how-to
│   └── onboarding-flow.task-type.md   ← proven pattern
└── src/
    └── ...

Each file suffix (.adr.md, .guide.md, .task-type.md) indicates its document type.

When an agent starts a session, Archcore’s MCP server makes all documents available. The agent can query, read, create, and link documents without leaving the conversation.

Why not flat instruction files?

Feature	Flat files	Archcore
Structure	Free-form text	Typed documents with templates
Scale	Gets unwieldy past ~50 lines	Separate documents, free-form directory structure
Reusability	Copy-paste between tools	One setup works with all agents
Findability	Agent reads the whole file	Agent queries by type, status, or topic
Lifecycle	No versioning beyond git diff	Draft → accepted → rejected status
Connections	No links between concepts	Explicit relations between documents

Read the full comparison →

What you get after setup

After running archcore init, your agents can:

• Find relevant context by querying decisions, rules, and patterns by type or topic.
• Follow team standards recorded as coding rules and how-to guides.
• Record new decisions and patterns as structured documents.
• Link related concepts, such as a plan to the PRD it implements, or a rule to the decision that produced it.
• Work from the same knowledge base across Claude Code, Cursor, Copilot, Gemini CLI, and more.

Two ways to use Archcore

• The Archcore plugin runs in Claude Code, Cursor, Codex CLI, and GitHub Copilot CLI. It adds four outcome-focused slash commands, built-in agents, and guardrails.
• The Archcore CLI is the core: one binary, 10 core MCP tools plus install_host_config for host wiring, and lifecycle hooks that work with all eight supported agents.

Both paths use the same .archcore/ directory. See Mental Model for how they fit together.

Next steps

• Mental Model explains the CLI as compiler and the plugin as runtime.
• Document Types covers the 21 types you can create.
• Use Cases shows the most common starting points for teams.