Skip to content
pmLet’s talk
All projects

Forge

AI media and agent orchestration in a macOS menu bar

Forge is a native macOS desktop app Pablo built as a personal AI orchestration center: one surface for generating media, running coding agents, and rendering demo videos. Instead of tab-hopping between provider dashboards and CLIs, everything runs through an Electron shell that spawns a local Express backend and streams progress to a React UI over WebSocket. It lives in the menu-bar tray and auto-launches hidden at login.

Forge: project preview
Role
Product design and engineering
Year
2026
Platform
Software project
Status
shipped

The problem

Producing AI media and demo videos means juggling provider dashboards, CLI tools, and one-off render scripts. Forge collapses that into a single macOS app: generate images, voice, and video, drive coding agents, and render Remotion demos from one window. Agents can reach the same media and repo tools over MCP.

What I built

The work includes electron/main.ts, server/index.ts, server/mcp-server.ts.

The choices behind it.

Solo · design + engineering

Implementation choice

AI-to-Remotion pipeline: a plain-English prompt becomes a JSON-schema-constrained Gemini component, written to a self-contained .tsx plus an isolated entry that registers only that one composition, then rendered with npx remotion render and streamed to the UI. The source project's Root.tsx is never mutated.

Implementation choice

Built-in MCP server (JSON-RPC over stdio) exposing 11 tools so Claude or any agent can drive Forge headlessly: generate image, video, and voice, list and read managed repos, and list, create, and browse templates and media. (Image edit, upscale, and template-run are declared but still stubbed on the MCP surface.)

Implementation choice

Four provider backends behind one surface: Gemini for image/video/speech/chat, ElevenLabs for voice, RunPod serverless GPU for Qwen-Edit / RealESRGAN / Qwen3-TTS, and the OpenAI Codex CLI run as a subprocess agent (codex exec --json).

Implementation choice

Per-purpose Gemini key routing: a pool of named API keys where image, video, audio, chat, and agentTasks each pin to their own key and model, resolved at call time.

Implementation choice

Live task and render streaming over WebSocket: agent output and Remotion render logs are pushed as they happen, each message tagged with its taskId so the UI can follow a specific run.

Implementation choice

Native macOS shell: a menu-bar tray app that spawns the Express backend as a child process via ELECTRON_RUN_AS_NODE and encrypts every API key at rest with Electron safeStorage instead of writing secrets to config.

MCP tools (JSON-RPC over stdio)
11
provider backends (Gemini, ElevenLabs, RunPod GPU, Codex CLI)
4
REST endpoints
60+
Build detail
12-section desktop UI

Under the surface.

electron/main.ts

app

macOS menu-bar shell that spawns the Express backend as a child process, holds the tray menu (show window, MCP status, copy LAN IP), and stores provider keys in the Keychain.

server/index.ts

service

Express 5 API on port 3400 that routes media generation, agent tasks, Remotion templates, git repos, providers, and workspace files, with a WebSocket for live task and render progress.

server/mcp-server.ts

mcp

Standalone stdio MCP server that hands Claude Code eleven forge_* tools (generate, edit, and upscale media; list repos and media; run and create templates) by calling the Forge API.

src (Forge dashboard)

app

React 19 and Vite dashboard (templates, images, video, audio, gallery, editor, agents, repos, providers, workspace, settings) that drives every backend route and is the localhost:3400 browser view.

server/gemini-compose.ts

library

Asks Gemini for structured JSON plus relevant SKILL.md context to synthesize a self-contained Remotion component, then writes it under the active project's src/ai-generated/ folder.

server/remotion-templates.ts

library

Remotion template engine that scans a project's Root.tsx for literal Composition tags, tracks parameterized and standalone templates, and maps rendered videos into the shared media root.

server/providers

library

Provider adapters for Gemini (Imagen, Veo, TTS), ElevenLabs voice, RunPod serverless GPU (Qwen-Edit, RealESRGAN, Qwen3-TTS), and Codex and Claude CLI subprocess launchers for agent tasks.

server/skills.ts

library

Skills bridge that indexes SKILL.md files across the Claude, Stitch, and Projects skill roots and loads raw content on demand for Gemini's system prompt.

Build details Stack, scope, and implementation

Built with

  • TypeScript 5.9
  • React 19
  • Electron 41
  • Vite 8
  • Express 5
  • Remotion
  • Google Gemini (@google/genai)
  • Tailwind 4
  • WebSocket (ws)
  • shadcn / base-ui
  • simple-git

Areas

  • AI orchestration
  • desktop app
  • MCP
  • media generation
  • Remotion
  • Electron

More about the project

The centerpiece is the AI-to-Remotion pipeline. Describe a video in plain English and Forge asks Gemini for a schema-constrained component, writes a self-contained .tsx with an isolated entry that registers only that one composition, runs remotion render, and streams the log back to the UI without ever touching the source project's Root.tsx. Forge also mirrors its powers as an MCP server, so Claude or any agent can generate images, video, and voice, read repos, and manage templates headlessly. API keys never hit disk in plaintext: they are encrypted at rest with Electron safeStorage, with env fallbacks for headless dev.

Try the product.

Interactive demo · Open full screen