Archcore Plugin — Spec-Driven Development & Context Engineering for AI Coding Agents

Make your AI coding agent work like it already knows your repo.

Archcore brings spec-driven development and automatic project context to Claude Code, Cursor, Codex CLI, and GitHub Copilot CLI. Specs, architecture, decisions, rules, and plans live in Git and are applied as the agent works.

The plugin pairs with Archcore CLI: the CLI provides the git-native context layer and MCP tools; the plugin adds skills, slash commands, computed routing over gated instruments, and guardrails.

Spec-driven development defines intent. Context engineering supplies the broader project understanding needed to execute that intent correctly — a spec is one part of context, not the whole context.

See it work

The agent pulls in the rules and decisions that apply — no command needed — and still gives you the four commands below for anything explicit.

Commands

Describe what you want in plain English — Archcore computes the route. The slash commands below are shortcuts to the same instruments.

Everyday context needs no command at all: hooks inject the applicable rules and specs when the agent edits a file, and the session starts with a recap of what's decided and in progress.

Inside the commands: the conductor and instruments

/archcore:plan computes its route instead of running one fixed flow. Grounding derives what the request changes in the recorded canon (the delta), where the missing information lives, how hardened the touched zone is, and which risk flags apply — then the conductor assembles the matching document package and announces it in one line. A layout fix takes the null route and produces no documents; one new capability gets a spec and a plan; a large initiative gets an umbrella PRD with one spec per capability. /archcore:document and /archcore:review route by wording into the same instrument set — decision, describe, actualize, closeout, experience:

You never pick a route or a size — the announcement names both before any document is created (naming an expert path still works: plan research, plan iso, review closeout). Every gate skips itself when an existing document already covers it, so a fully-specified request runs question-free; a vague one stays within 5 questions. An interrupted flow resumes in a later session — the draft document carries the route state.

plan research <topic> runs the research instrument, which produces either a research document (a territory survey closed by scope coverage) or an rnd (an investigation closed by a recommendation, selected when the request names a pending decision or a set of candidates); document research <report> files a ready report the same way; document evidence <material> files one external material as an evidence document. The research and evidence types and the supports, contradicts, and supersedes relations require Archcore CLI ≥ v0.8.3; on an older CLI, plan research falls back to an rnd and reports the required version.

Install

Archcore plugins require the Archcore CLI on PATH — it serves the MCP server the plugin talks to.

# macOS / Linux / WSL
curl -fsSL https://archcore.ai/install.sh | bash

# Windows (PowerShell 5.1+)
irm https://archcore.ai/install.ps1 | iex

Verify: archcore --version · Update: archcore update · Docs: docs.archcore.ai/cli/install

Then add the plugin in your host:

Claude Code

/plugin marketplace add archcore-ai/plugin
/plugin install archcore@archcore-plugins

Cursor — requires Cursor 2.5+. Open Plugins, paste https://github.com/archcore-ai/plugin into Search or paste link, click Add Plugin. One-time MCP setup: copy docs/cursor.mcp.example.json into ~/.cursor/mcp.json (user-scoped) or .cursor/mcp.json (project-scoped).

Codex CLI — requires Codex CLI v0.117.0+.

codex plugin marketplace add archcore-ai/plugin
codex
# then run /plugins, open Archcore, select Install plugin

GitHub Copilot CLI — two steps, both required. The plugin brings skills, commands, agents and hooks — but on Copilot it cannot bring the MCP server, so a project that skips step 2 has no document tools at all.

# 1. Install the plugin (from this repo's plugin subdirectory)
copilot plugin install archcore-ai/plugin:plugins/archcore

# 2. Wire your project (registers the MCP server; run once per repo, commit the result)
archcore init --agent copilot --project "$PWD"

Copilot CLI only: VS Code agent mode has no self-serve plugin install, and cloud-agent sandboxes do not load plugin hooks.

Why step 2 is not optional: a plugin's MCP server is launched in the plugin install directory with no project path (github/copilot-cli#4234), so it would serve the plugin cache rather than your repo. The plugin therefore ships no MCP server to Copilot at all, and the project-level one from step 2 is the only source of document tools. Archcore CLI ≥ v0.6.7 refuses to serve from a plugin cache by design, so the failure mode is loud rather than silent.

claude --plugin-dir /path/to/plugin

ln -s /path/to/plugin ~/.cursor/plugins/local/archcore
# then in Cursor: Cmd/Ctrl+Shift+P → "Developer: Reload Window"

codex plugin marketplace add /path/to/plugin
codex plugin add archcore@archcore-plugins

copilot --plugin-dir /path/to/plugin/plugins/archcore

Try these first

Open your project and try these three prompts. Each shows a different side of what your agent can now do.

Empty repo? Run /archcore:init first — it seeds a stack rule, a run-the-app guide, and optionally imports your existing CLAUDE.md / AGENTS.md / .cursorrules.

1. "Before I change anything in src/auth/, what should I know?" Your agent sees what's already decided for that path — before it touches the code.

2. "Add a new API handler and follow this repo's conventions." Your agent places the handler where your architecture says it belongs, instead of guessing.

3. "We picked PostgreSQL — record it as a team standard." The decision is captured, codified as a rule, and auto-applied to every future change in the same area. Decisions stop dying in chat scrollback.

What changes after install

❌ Without Archcore

• the agent guesses your folder structure
• re-litigates decisions your team already made
• needs the same conventions repeated in every chat
• loses project truth the moment the session ends

✅ With Archcore

• code lands where your architecture says it belongs
• respects decisions already in Git
• follows team conventions loaded automatically
• reflects new decisions as future guardrails — not markdown graveyards

The agent stops guessing and starts following the system.

Use Archcore when

• Your agent writes code, but not the way this repo expects
• Your CLAUDE.md / .cursorrules / AGENTS.md keeps growing and drifting
• You work with 2+ agents or 2+ host tools (Claude Code + Cursor + Codex + Copilot)
• You want decisions, rules, and specs in Git — not in chat scrollback

Not for — chat memory, a prompt library, or a one-shot spec-to-code generator. Archcore is a git-native context layer for AI coding agents, not a methodology kit.

Supported hosts

Built on open standards (Agent Skills, MCP) — skills and MCP tools are shared across hosts; only manifests are host-specific.

How Archcore differs

Pick a methodology tool for an opinionated dev flow. Pick a memory tool for session continuity. Pick Archcore when you want typed, queryable project truth that your coding agent respects on every request.

Uninstall

Claude Code: /plugin uninstall archcore@archcore-plugins Cursor: remove from plugin settings. Codex CLI: codex plugin uninstall archcore GitHub Copilot CLI: copilot plugin uninstall archcore

License & contributing

Apache-2.0 · Issues and ideas: GitHub Issues