Autonomous Video Studios
HomePART III · ChannelForge
CHAPTER 14 · CHANNELFORGE

Architecture

ChannelForge is a web application, a control plane and a GPU fleet. The web application is what people and programs talk to; the control plane does the work; the fleet renders. They meet in a PostgreSQL database and a Redis instance.

Members in the browser

workspaces · teams · roles

REST API · webhooks

API keys, signed deliveries

MCP server · n8n · Make · Zapier

agents drive it as tools

Cloudflare › nginx › two Next.js instances (rolling deploys)
forge-standby · Google Cloud · europe-west1
App (Next.js 15, TypeScript)

pages, actions, OAuth connects, billing portal, admin console

PostgreSQL 16

Cloud SQL, Auth Proxy, nightly + 7-day PITR

Redis

queues, budgets, rate limits, progress

Control plane — Python services under systemd; the model gateway is Node.js
workers
cpu · io · film queues
scheduler
slots, autopilots, buffers
fleet
GPU boxes and their slots
sentinel
reaping, prunes, alerts, caps
gateway
model metering and budgets
voice
Chirp 3 HD narration
mailer
e-mail through Amazon SES
billing
Stripe · PayPal
AWS GPU fleet

its own L4 boxes (1a, 1c), 2 slots each, sandboxed

Amazon Bedrock

Sonnet 5.5 · Opus 5.5 · Fable 5.1/5.2, via the gateway

Google Cloud Storage

tenant media (S3 API), private, per workspace

Eleven platforms

YouTube · Meta · LinkedIn · X TikTok · Threads · Bluesky…

Figure 29. ChannelForge's architecture.

The web application#

A Next.js 15 application in TypeScript (React 19 server components, server actions, Tailwind CSS) serves the marketing site, the documentation (82 pages), the signed-in product and the admin console (42 staff pages). Two instances run behind nginx, which retries a request on the other instance, so a deploy rolls one instance at a time; a release builds first and stops the control units around its migrations. Prisma maps the database's 313 tables.

The control plane#

ServiceRole
worker-cpu / worker-io / worker-filmjob queues by resource: planning and copy; uploads and platform calls; film plans and audits
schedulerslot decisions, autopilots, keep-ready buffers, generation per day
fleetstarts and stops GPU boxes, assigns jobs to slots, sets boxes up, installs code
sentinelreaps dead jobs, prunes, alerts, enforces spending caps, runs privacy purges
gatewaya Node.js service, the only door to the models: it speaks the Anthropic Messages API, holds the one Bedrock credential, meters every call per workspace and job, and enforces budgets before a call
voicethe narration service (Chirp 3 HD) the GPU jobs call through a tunnel
mailertransactional email through Amazon SES, deduplicated alerts
billingStripe events and PayPal resync, plan credits, invoices, disputes

Every unit of work is a job with a kind, a resource, a state, attempts and an event log the customer can watch live. 2,735 jobs have run so far, of 47 kinds, from niche suggestions and channel drafts to episode generation, translations, films, insights and comment replies.

Data#

PostgreSQL 16 holds workspaces, users, teams, channels, episodes, posts, jobs, credits, plans, invoices, insights and the audit log — 313 tables. It runs on Google Cloud SQL, reached through the Cloud SQL Auth Proxy, with nightly backups and seven days of point-in-time recovery. Redis carries caches, budgets, rate limits and live progress. Media lives in a private Google Cloud Storage bucket, under one prefix per workspace.