DropIn
A video workspace for agencies and brand content teams. One finished video goes from raw footage to client approval to a live post on YouTube, Instagram, Facebook, LinkedIn and TikTok, from one shared calendar. I designed and built it solo.
The problem
A team makes one good video and then has to ship it five times. Every platform has its own rules for length, shape, size and captions. Approvals happen over email, copy gets rewritten by hand for each platform, and a post that broke a rule is only discovered after it misses its slot.
DropIn puts the whole path in one place and checks every destination before anything is scheduled.
What a team gets
- One place from footage to live post, with boards, a shared calendar and status that matches what's really on each platform.
- Client approvals with a record. Every decision is logged, so there's no argument about who approved what.
- Copy drafted per platform. Claude writes titles, captions and tags from what was said in the video.
- Failures caught early. A post that would break a platform's rules is blocked before it's scheduled. The rest still go out.
- Room for agents. AI agents can plan and schedule through the same permissions a person has.
How a video moves
- Footage uploads to private storage and is transcribed in the background.
- The team picks destinations. Claude drafts titles, captions and tags for each one.
- A readiness check tests each destination against that platform's rules.
- The team or client approves. Every decision is logged.
- A scheduled job publishes at the slot, with retries.
- Live links and status flow back into DropIn.
By the numbers
5 apps, 13 data domains, 16 background workers, 99 migrations, 711 commits from March to July 2026.
Click any box for what it does, why it matters, how it works and what it cost. Filter by skill, or press a ▶ walkthrough.
Where AI does the work
| Job | Model | Why this way |
|---|---|---|
| Titles, captions and tags per platform | Claude | It writes from the transcript to each surface's limits, so a Short and a LinkedIn post get different copy from one upload. Cost per call is capped by plan limits. |
| Transcripts | Whisper | Runs once per file and is cached, so no video is ever paid for twice. |
| Agent access | MCP server, OAuth 2.1 | Outside agents can upload, draft, check and schedule with the same permissions as a person. |
Guardrails
- AI drafts, the readiness check decides. Generated copy still has to pass the same platform rules as anything a person writes.
- Agents get no back door. MCP calls pass the same auth, organization, role and workspace checks as the web app, with short-lived tokens.
- AI work runs off the request path. Transcription and copy generation are queued jobs, so a slow model call never blocks the app.
Stack
| Layer | What |
|---|---|
| Monorepo | Turborepo + pnpm, TypeScript strict. Apps: web, api, marketing, docs, e2e. Shared packages: types, config, db. |
| Frontend | Next.js 16, React 19, Tailwind 4, Catalyst UI, TipTap, dnd-kit. Marketing and help docs on Astro. |
| API | Fastify 5, Zod 4, OpenAPI with Scalar docs, BetterAuth (organizations, roles, MCP plugin). |
| Data | PostgreSQL 18 with Drizzle, Redis, BullMQ with 16 workers, Cloudflare R2 for media. |
| Cloud | AWS CDK in TypeScript: ECS Fargate behind a load balancer, RDS Postgres, ElastiCache Redis, CloudFront + S3 for marketing. Separate AWS accounts for staging and production. |
| CI/CD | GitHub Actions with OIDC, so no long-lived cloud keys. Migrations run in a Fargate task before each deploy. |
| Testing | Playwright end-to-end against real Postgres, Redis and MinIO in Docker. |
Key decisions
- Destination = platform + surface. An Instagram Post can't take a video. Modeling surfaces stops a 12-minute cut from going to a Reel before anyone clicks Publish.
- One rulebook. The surface matrix lives in one file. The UI, the readiness check and the publish worker all read it, so they can't disagree.
- Adapters at the edge. Validation never learns vendor vocabulary. Moving Facebook and Instagram to Meta's own API touched the adapter and nothing else.
- Direct where it counts. YouTube and Meta go direct for reach and analytics. LinkedIn and TikTok go through an aggregator because their APIs are slow to approve.
- Don't trust webhooks. They get lost or replayed, so polling and a nightly reconcile keep status honest.
- Hand-written migrations. Every schema change ships with an idempotent migration in the same commit. A failed migration stops the deploy.
- R2 for video. S3-compatible with no egress fees, and platforms pull the final file by URL.