Castforge

Castforge is a Windows desktop application built by Julien Abenhaim, a developer in Saint-Bruno-de-Montarville, Quebec, and operated by 9569-2893 Québec inc. It opened as a public beta in September 2026 and is currently free during early access. The tool lets developers connect the AI subscriptions they already pay for — GitHub Copilot, Claude, OpenAI Codex, Google Gemini, and others — and run them as a coordinated team on a live project board.

What It Is

Castforge is a multi-agent AI coding workspace in the desktop app category. Rather than running a single AI assistant, it assigns distinct roles to different agents: a Lead who plans, Coders who build, a Tester who runs checks, a Reviewer who validates the work, a Researcher, and a Designer who guards the UI against a design spec. These agents hand off to each other, run review loops, and work in parallel on isolated branches. The developer watches and steers from a live course board with a live preview of the app being built.

How the Team Model Works

Each project is treated as a staffed team with enforced role boundaries. The Lead drafts a phase board automatically; cards move as work progresses, and dependencies are enforced so nothing starts before the work it needs is done. Agents can run in parallel, each on its own git worktree, and merge through a review gate. A Designer role reviews the UI against a design spec rather than just the code. Permission modes — Careful, Balanced, or Autonomous — let the developer control how much the team acts without approval.

Key team mechanics include:

• Presets (Solo, Pair, Trio, or custom) to configure the team quickly
• Adaptive intent routing that matches process depth to request size — a one-line fix runs immediately, a large feature gets a full plan
• Quality tiers (Draft, Standard, Max) that control which review gates run, including a plan-check gate before any code is written and a goal-backward verification gate on finished work
• Self-healing runs that name and repair a card that reaches Done with no record of where its work was written, then report the repair rather than going quiet

Bring Your Own AI

Castforge does not sell model access. Developers connect up to 17 AI coding agents using accounts and subscriptions they already have. Five agents sign in with an existing account: GitHub Copilot (free), Claude, OpenAI Codex, Google Gemini, and Cursor. Local models require no account. OpenRouter and Factory Droid are also supported. Each provider bills the developer directly; Castforge never marks up model usage.

Project Memory and the Decision Ledger

Castforge maintains a living project state — stack, what is built and wired, key decisions, invariants, and current focus — that every fresh agent reads before touching the work. This state is backed by an append-only ledger filterable by decisions, criteria, gate outcomes, and transitions. The ledger can be exported to a decisions.md file that commits alongside the code. The Time Machine feature lets developers scrub back through snapshots of the app and resume from any earlier point.

Platform and Current Status

The current release targets Windows 10 and 11. macOS and Linux builds are in active development and listed on the public roadmap. A mobile companion for approving agent prompts remotely is also planned. Web publishing (Vercel URL and custom domain) is rolling out across beta builds; iOS, Android, and desktop publishing targets are listed as coming soon.

The tool is in open beta at version v0.6.7-beta as of the time of writing. According to the about page, most changelog entries trace back to reports sent from inside the app or posted in the Discord community. The founder states that the project started as a personal tool in spring 2026 before opening to public beta in September 2026.

Integrations and Backend Scaffolding

Castforge connects to GitHub, Vercel, and Supabase as tools agents can use directly. It also provides guided setup for common backend pieces — auth, database schema and migrations, payments (Stripe), transactional email, file storage, scheduled jobs, and a test harness — framed as short guided setups rather than boilerplate the developer writes manually.