Changelog

New preset in the video generator: UGC. The "iPhone in someone's hand talking about a product" look, without the actual person or the actual iPhone.

What it wires up under the hood:

• Prompt scaffolding. The router gets an extra system message steering it toward handheld framing, natural indoor lighting, medium close-up, no cinematic camera moves. Aspect ratio locked to 9:16.
• Burned-in captions. Post-render we run an FFmpeg pass with subtitles=captions.srt:force_style='Fontname=Inter,Fontsize=18,PrimaryColour=&HFFFFFF&,OutlineColour=&H000000&,BorderStyle=3,Outline=1' — the TikTok-native look, not a subtitle track a platform might strip.
• Humanized TTS. ElevenLabs eleven_multilingual_v2 with stability=0.35, similarity_boost=0.75. Lower stability on purpose — the tiny imperfections read as human.
• Product photo upload. Drop a JPEG/PNG in the UGC panel and it gets passed to Runway as a reference image so the product shape stays consistent shot-to-shot. Nothing fancy; the reference-image field was already there, we just exposed it.

Meant for creator-brand ads. Works surprisingly well for SaaS demo cutdowns too.

Captions used to sound like a robot doing an impression of you. Now they sound like you.

• Voice profiles live in Postgres with a pgvector embedding column. Paste in 10–50 sample captions, we chunk and embed with text-embedding-3-small (1536-dim), and store them per workspace.
• RAG at caption time: when you hit "generate caption", we pull the top-k nearest neighbors to the current draft topic (cosine distance, k=6) and inject them as few-shot examples. The model gets your actual phrasing, not a description of it.
• Dashboard at /dashboard/voice: create profiles, paste samples, mark a profile as default, preview how a caption comes out with vs. without the profile.

Tested it against my own last 200 TikTok captions — the with-profile output is close enough that I don't need to rewrite it. That was the whole bar.

You can now drive the whole app from Claude Desktop, Cursor, or n8n's AI Agent node without touching the UI.

• /api/mcp — Model Context Protocol endpoint. Streamable HTTP transport. Six tools out of the gate: list_posts, create_post, schedule_post, generate_caption, upload_media, get_post_status. Auth via Authorization: Bearer pk_live_....
• /api/docs — OpenAPI 3.1 spec plus a Scalar-rendered reference. Same auth model. Every dashboard action has a matching endpoint; if it's not in the spec, it doesn't exist.
• API keys live under Dashboard → Settings → API keys. Hashed with SHA-256 before insert; we show the plaintext exactly once at creation. Scoped per-workspace, revocable, last-used timestamp is tracked so you can spot stale ones.

The reason this exists: I wanted my own agents to schedule my own content. If you want the same, you now can.

First public build is up at poster.abmediax.com. Been running it privately for weeks; time to let other people use it.

What works today:

• YouTube posting is live and reviewed. Uploads via resumable session, sets title/description/tags, honors your default privacy setting.
• TikTok, Instagram, X are wired end-to-end but pending platform review. You can connect the accounts now; publishing turns on the moment each app is approved. No code change needed on your end when it flips.
• Captions via OpenAI (gpt-4o-mini for drafts, gpt-4o for the "make it better" pass). Prompt is boring on purpose — hook, body, CTA, hashtags — and it respects the brand voice profile if you have one set.
• Auth is Supabase (email + Google). Sessions via @supabase/ssr, middleware-gated everywhere except /, /login, /changelog, /api/docs, /api/mcp.
• Storage is Vercel Blob. Videos land there first, then get pushed to each platform. No S3 bill to reason about.
• Custom domain: poster.abmediax.com, HTTPS via Vercel, wildcard subdomain reserved for future workspaces.
• Pro plan ($19/mo) unlocks the scheduler (Vercel Cron every 5 minutes checks posts where status='scheduled' AND scheduled_at <= now()).
• OAuth tokens for each connected platform are encrypted at rest (AES-256-GCM, key from TOKEN_ENCRYPTION_KEY env). We never store platform passwords, only refresh tokens.

There will be bugs. Report them and I'll ship a fix the same day.

Shipped a one-box AI video generator. Type any prompt, get a finished clip in your workspace. Preview-only — nothing auto-posts until you approve it.

How it works:

1. Router. Your prompt goes to a gpt-4o router that plans cinematic shots — subject, camera move, lighting, duration per shot. Cheap and fast; output is a strict JSON schema.
2. Render. Shots go to Runway Gen-4.5 as the primary provider. There's a Sora fallback stub wired up but disabled until API access lands.
3. Quality gate. Every returned file is probed with ffprobe. If duration is off by more than ±0.5s, resolution is below the requested tier, or audio track length doesn't match video length, we reject and retry once. Failed twice = surfaced to you with the probe output so you know why.
4. Delivery. Passing clips land in Vercel Blob and show up under /dashboard/video as previewable assets. You pick one, edit the caption, then send it to the poster.

New route: /dashboard/video. Runway credits pass through at cost — no markup, no re-billing.