JornAI

From source to Instagram feed in minutes — AI writes the copy, a human always approves.

An open-source web tool that lets a newsroom publish breaking-news posts to Instagram with almost no friction: paste a police report, a press release or a link, and get a ready-to-review post with a title, subtitle and caption — built on a pipeline designed not to invent facts.

_{Every story in one feed: paste a source → AI drafts the post → frame the photo → a human approves and publishes.}

TL;DR: paste a text, an article link or a PDF, optionally credit someone (📸 @photographer), and hit Generate. The AI writes the title, subtitle and caption — and never touches the photo.

TL;DR: drag and zoom the photo, drag the title or subtitle to reposition them, edit the words (69 / 149 character limits). The brand template stays intact; only this post changes. What you see is exactly what the server renders.

Video (Reels) works too. Upload a clip instead of a photo and JornAI handles the boring parts:

• Normalizes any video to 9:16 — already-vertical clips are cropped top and bottom; landscape clips are kept whole over a blurred, mirrored backdrop instead of ugly black bars.
• Keeps the title card and logo inside Instagram's safe zones, so the profile, caption and buttons never cover them.
• Burns in an animated title card (slide + fade) in one of three styles, using your brand colors.
• Fixes phone-recorded videos that arrive rotated.

TL;DR: add up to 10 photos and the post becomes an Instagram carousel. The 1st photo is the cover, with the template, title and subtitle; the others go out as the bare photo, cropped to the same proportion (Instagram requires it). Reorder, swap the cover and frame each slide.

TL;DR: every photo — cover or slide — has an Edit image popup, framed at the real proportion of the art:

• Blur image — tap or drag over the photo to hide faces, plates or anything else (with undo). Handy for accident and police photos.
• Adjust — zoom, straighten, rotate, mirror, and blurred borders: show the whole photo and fill the rest of the frame with a mirrored, blurred copy, the same look as the videos.
• Color — brightness, contrast and saturation.

TL;DR: see the post exactly as it will look on Instagram (carousels can be swiped), then Approve and publish, Request rewrite (the AI writes a new version) or Reject with a reason. Every change is a new version, so the full history is kept.

• Add to story too — a checkbox under the approve button, always unticked by default. When ticked, once the feed post is out the art (carousel: cover only) or the rendered video is also published as an Instagram story.
• If the story fails, the feed post stays up — the failure is recorded and the review screen says why.

TL;DR: the admin area is where a newsroom makes JornAI sound and look like itself.

• Newsroom style — register real posts from your newsroom; the AI learns the tone and format (it never copies the content).
• Templates — build photo templates visually (see below).
• Users & roles — admin, manager and reporter, each company isolated from the others.
• API keys, Company, Settings, AI balance — connect providers, set your logo and brand colors, toggle mandatory review, watch your AI spend.

TL;DR: the photo template is the heart of JornAI, and it is fully customizable — upload your brand frame (a transparent PNG made in Canva or Photoshop), drag the title and subtitle boxes where you want them, and pick the format (4:5 or 1:1), font size, color, weight, alignment and case. Video is simpler on purpose: it ships with 3 ready-made title-card styles (classic, light, bold) that pick up your brand colors.

Newsrooms often need to get a short, urgent post out — an accident, a weather alert, a local event — with almost no friction, but still with editorial review and consistent branding. JornAI is not a generic social-media scheduler and not a raw AI text generator with no guardrails:

• 🧾 Fact-first. Sources are cleaned, redacted and checked before and after the model, so a confident-sounding article doesn't come with invented details.
• 👀 Always human-approved. Nothing reaches Instagram without a person approving it.
• 🎨 Consistent branding. Photo + text are composed on a fixed template, in the browser and on the server, with identical results.
• 🌍 Works in your language. One environment variable picks the language of the sources and the posts (en and pt included, more welcome).
• 🆓 Runs with no keys. AI_PROVIDER="mock" + local storage lets you try the whole flow without any external service.

Source (photo, text, link, or PDF)
        │
        ▼
AI writes the copy (title, subtitle, caption)
        │
        ▼
Reporter frames the photo(s) or video in the brand template
(single photo, carousel of up to 10, or a 9:16 Reels video)
        │
        ▼
Editor reviews → approve / reject / regenerate / edit
(optional: "Add to story too")
        │
        ▼
Publish to Instagram (feed / carousel / Reels, + story if ticked)

• AI only writes text — it never generates or picks images. Photos are always real and chosen by a person.
• Every AI generation (or manual edit) creates a new version, so nothing is overwritten and there is a full history.
• Photo, carousel (up to 10 photos) and video (Reels) posts, with an animated title card for video.
• Built-in image editor: blur faces, blurred borders, rotate/straighten, color.
• Optional story share on approval — always an explicit, per-post choice.
• Peer review: a reporter's post can be approved by a manager/admin, or by two colleagues.
• Multi-tenant: each company has its own feed, templates, style examples and settings.

Feeding a police report or an official PDF straight into an LLM is a good way to get a confident-sounding article with wrong facts. JornAI's pipeline is built around not letting that happen:

• Source text is cleaned and compacted deterministically (no AI) before it ever reaches a model — official documents/forms get their real fields extracted (who, where, cause, injuries) instead of being blindly truncated.
• Personal data (names, national IDs, phone numbers, plates) is redacted before the AI ever sees it.
• The system prompt has explicit, tested rules against inventing details, generic filler, or unsupported claims.
• Generated content is checked against the source by a deterministic validator after generation (e.g. the weekday must match the source date, no unsupported claims about investigations, road closures or deaths); a violation triggers one corrective regeneration before failing loudly.
• The weekday is computed by code, never by the model.

JornAI works in any language — one per deployment: the language of the sources and of the generated posts, chosen with the APP_LANGUAGE environment variable. Each language is a small pack with its own rules for source cleanup, personal-data redaction and the fact validator. Two ship today, as examples:

The prompts themselves are written in English; only the examples and rules the model must reproduce live in the language pack (src/lib/language/<lang>/). To add your language, add a pack there and register it in src/lib/language/index.ts — see CONTRIBUTING.md. APP_LANGUAGE also sets the default interface language (each person can still switch it in the UI); the code, comments and logs are always English.

• Next.js (App Router) + TypeScript, PostgreSQL via Prisma
• Fabric.js for the in-browser art editor, Sharp for the server-side final render, ffmpeg for video
• AI: any OpenAI-compatible or Anthropic text provider (OpenRouter, Groq, Gemini, OpenAI, Claude), with automatic fallback across providers
• Instagram Graph API for publishing (single image, carousel, Reels and stories); S3/R2-compatible storage (or local disk in dev)

git clone https://github.com/MatheusBomtempo/JornAI.git
cd JornAI
npm install
docker compose up -d       # PostgreSQL that matches .env.example (or bring your own)
cp .env.example .env       # fill in DATABASE_URL and, optionally, an AI provider key
npm run db:migrate
npm run db:seed            # sample company + accounts, dev only
npm run dev

Open http://localhost:3000 and sign in with admin@jornai.local / admin12345 (created by the seed, development only).

No AI key yet? Keep AI_PROVIDER="mock" in .env to run the whole flow with a stub AI response — useful for trying the app without any external service.

Two kinds of configuration, done once, in this order:

Everything technical is configured through environment variables — see .env.example, which documents each one. The important ones:

TL;DR: before the first post, open Admin and make JornAI look and sound like your newsroom. It takes a few minutes and you only do it once.

• Company — your name, logo, Instagram @ and brand colors (the logo and colors show up on the art and on the video title cards). On the very first login the admin is guided through this.
• Templates — build your photo frame with the template builder: 4:5 or 1:1, your overlay PNG, title and subtitle positions and text style.
• Newsroom style — paste a few real posts so the AI learns your tone and format.

You can come back and change any of these later.

The project deploys as a regular Next.js app (it is developed against Vercel + Neon + Cloudflare R2). npm run build applies the Prisma migrations, so the build needs the database configured.

• SPEC.md — full technical spec: database schema, API endpoints, state machine, Instagram publishing.
• CONTRIBUTING.md — how to contribute and how to add a language.
• SECURITY.md — how to report a vulnerability.

• More language packs (Spanish, French, …)
• More official-document profiles (the way police-report forms are cleaned today)
• Other social networks (deliberately out of scope for now)

Not planned, on purpose: scheduling, AI image generation, and publishing without human approval.

Issues and pull requests are welcome — read CONTRIBUTING.md first. If JornAI is useful to you, a ⭐ helps other newsrooms find it.

MIT — see LICENSE.