Reactor
From a design to tested software. With a team of agents you review.
A self-hosted AI studio for software houses: Figma design, epics and stories, a breakdown into tasks per repository, Claude Code and Codex implementing them side by side, a live container preview, Playwright tests that file their own bugs and a wiki for people. Plus an agentic IDE that keeps the backlog written for you.
Get started · Website · Architecture · Report a bug · Čeština
Get started
You need Node.js 22+, Docker (Docker Desktop or OrbStack) for the preview and tests, and a signed-in Claude Code (claude); Codex (codex) is optional.
git clone https://github.com/DanielHromadka/aforge.git
cd aforge
npm install
npm run runtime:build # once: the aforge-runtime image (Ubuntu 24.04, Node 22, .NET 8, Python 3, Nginx)
AFORGE_AUTH=off npm run dev # server :4310 + web http://localhost:5310, one local administrator
With sign-in: npm run keycloak starts Keycloak with the aforge realm on :8380, then npm run dev and sign in as admin / admin or developer / developer (local development accounts only). Production on one Linux server, with Cloudflare Tunnel, backups and sandboxed agents: deploy/production/INSTALL.md.
The UI is in Czech. The complete feature reference is README.cs.md.
Features
Work flows by story
- Jira-like work: epic → story → task, bug and sub-task, with keys like
KV-7 in one numbering. A header ladder of steps (Overview, Design, Stories, Breakdown, Implementation, Preview, Tests, Wiki) that is also the navigation, one colour and icon per step everywhere.
- Design from Figma: frames become screens and layers become elements; notes, linked stories and a being worked on flag live beside them.
- Analyst proposes epics and stories with Given/When/Then criteria; Analyse and reorganise reviews a backlog before implementation (merge, split, move, clarify) with every original criterion accounted for.
- Architect breaks stories into frontend, backend and test tasks for the right repository and keeps one OpenAPI contract written into every repository.
- One agent per repository, repositories in parallel: follow the run as a timeline, open the changed files side by side, then accept or return the task with a comment.
- Preview in an isolated container with Lovable-style annotations: a note on an element of the running app is a bug.
- Playwright tests map to stories by
[KEY-n] in their title; a failing test opens a bug and passing again closes it; walkthroughs are recorded as video.
- Wiki for people: after green tests the Documenter proposes what each feature does, how, and its rules.
- Planning for people: estimates, notes, reminders and deadlines that no agent ever sees.
The agentic IDE
- Composer (⌘L) plans a prompt into tasks per repository in one call and starts their agents; follow-ups queue behind the running run. Ask mode is the consultant.
- Editor (Monaco) with the agent's inline diff, accept or return in place, and ⌘K to send a selection straight to its repository's agent.
- The Scribe writes the backlog from what you ask: a new topic becomes a story (and an epic), a follow-up adds criteria, a technical change stays a task, and an unsure prompt waits unsorted instead of being guessed.
- The trace (⌘⇧T) shows what the Scribe and the Documenter wrote from every prompt; every AUTO record can be reverted, and the board marks such stories.
- Ponor (⌘⇧↩), the immersive mode: only the running app and one prompt line; click an element to talk about it.
- Switch any session between Studio, IDE and Vibe coding at any time; every story, task and prompt stays.
Agents on your own servers
- Workers run where Claude Code or Codex is signed in on a subscription, connect out to a relay (no open ports) and exchange end-to-end encrypted messages (ChaCha20-Poly1305, one key per worker).
- Pipeline: an orchestrator plans the work by phases with dependencies; the hub hands independent tasks to free workers by capability, label and load.
- Runs and usage: one list of local and remote runs with model, worker, step, duration, context and reported tokens and cost.
You stay in charge
- Agents propose, people decide: consultant answers are proposals to accept or reject, and code always waits for review.
- Everything has a way back: git commits, versioned descriptions and wiki pages, revertible Scribe records, sessions and read-only snapshots.
- Previews, tests and production agents run in bounded containers with filtered outbound networking, never with the host's environment or Docker socket.
- Projects are plain JSON and git repositories in
data/, on your own server, with Keycloak roles (owner, editor, viewer).
Building
npm run typecheck # the main gate: shared, protocol, runner, server, relay, worker, web
npm test # vitest
npm run build # the web app; npm start -w @aforge/server serves it on :4310
npm run release # the tested deployable artifact in .releases/
| Path | What it is |
|---|
packages/shared | the domain model and the zod schemas of the agent contract |
packages/protocol, packages/runner | the hub ⇄ relay ⇄ worker wire format; headless Claude Code / Codex runs |
apps/server | Fastify API, agents, builds, preview, tests, the IDE's Scribe |
apps/web | the React editor |
apps/relay, apps/worker | the relay and the aforge-worker daemon |
website/ | the website, published by GitHub Pages |
More in docs/ARCHITECTURE.md and CLAUDE.md.
Security
Please report vulnerabilities privately, see SECURITY.md.
License
GPL-3.0-or-later.
