Getting started with forge

forge is a single CLI that installs engineering-quality plugins into four AI coding assistants — GitHub Copilot, Claude Code, Codex, and Gemini CLI — from one authored definition. Your assistant gains the ability to analyze discussions, triage issues, review PRs, and coach your engineering practice, without forge ever running a background service or touching your data.

The package is on npm as forge-ai-assist and lives at arcanelabsio/forge on GitHub.

Install

npx forge-ai-assist@latest

That's the whole install. Forge asks which assistants you use, writes their plugin files, and gets out of the way. Requires Node 22 or newer.

To install everything non-interactively (useful in CI or a dotfile bootstrap):

npx forge-ai-assist@latest --assistants all --plugins all

Check what's currently installed:

npx forge-ai-assist@latest status

Uninstall cleanly:

npx forge-ai-assist@latest --uninstall

What you get

Forge ships three plugin groups. The Core group installs by default; the others are opt-in.

Core — read-only GitHub analysis, useful every day.

• Discussion Analyzer — "Summarize the top 5 open discussions and highlight unanswered questions."
• Issue Analyzer — "Show me open bugs labeled P0 and group them by component."
• PR Comments Analyzer — "Analyze review comments on PR #42 — what are the recurring themes?"

Elevate (--plugins elevate) — coaching on your engineering practice.

• Commit Craft Coach — "Review my last 10 commits — are they atomic and well-narrated?"
• PR Architect — "Evaluate my open PR — is it structured for easy review?"
• Review Quality Coach — "Assess my recent reviews — are they specific and architecturally deep?"

Ops (--plugins ops) — release machinery.

• Release Notes Generator — "Generate release notes from v1.1.20 to HEAD."

How you invoke a plugin

Each assistant surfaces forge plugins differently because each assistant has its own convention. Forge authors the plugin once and its adapters render the right shape for each tool:

Assistant	How to invoke
Claude Code	/forge:discussion-analyzer, /forge:issue-analyzer, …
Codex	$forge-discussion-analyzer, $forge-issue-analyzer, …
GitHub Copilot	/agent → pick a forge-* agent
Gemini CLI	forge:discussion-analyzer, forge:issue-analyzer, …

No matter which assistant you're in, the natural-language examples above work verbatim — you pass intent, the plugin handles the gh/git commands.

Selective installs

Install only Core plugins for only one assistant:

npx forge-ai-assist@latest --assistants claude --plugins core

Install the Elevate coaching plugins across all your assistants:

npx forge-ai-assist@latest --plugins elevate

Everything is additive. Re-running forge with a different set replaces only its own managed files; your assistants' other configuration is untouched.

How it works

Forge does not run a background service. It does not proxy your data. It does not cache answers or mint its own tokens. When you invoke a plugin, your assistant executes read-only gh and git commands in your current repo and interprets the results. The only thing forge ships is a set of prompt files — agents, skills, workflows — in each assistant's config directory.

The upside: zero trust surface added. The data your AI sees is exactly what gh already sees on your machine.

The consequence: forge plugins are strictly read-only. They can analyze, summarize, and coach, but they never mutate a repository or create a PR or comment on an issue. That boundary is intentional and enforced at the plugin-definition level.

What's next

• Package on npm — latest version and release history.
• Source on GitHub — full README, per-assistant setup guides, and the plugin architecture doc.
• Usage Guide — file locations, custom instructions, and tips per assistant.
• Plugin Architecture — how a single definition renders to four assistants.

More guides soon — a deeper look at the Review Quality Coach and the PR Architect, and a walkthrough of authoring your own forge plugin for internal use.