The short answer: yes. The Instagram Graph API has supported programmatic feed, reel, carousel and story publishing for Business and Creator accounts for years. The interesting question isn't can it · it's should the same system that publishes also decide what to publish. That's where things get real.
What Instagram allows, exactly
Any account converted to Business or Creator, connected to a Facebook Page, can authenticate an app to publish on its behalf through Meta's Graph API. Feed posts, reels, carousels and stories are all supported. The rate limits are generous for organic content, comment reads are effectively unlimited, and reply-to-comment endpoints work end-to-end. Meta doesn't care whether the "person" hitting POST is a human clicking a button or a job runner triggered by a cron · it cares that the token is valid and the account is in good standing.
The two-stage pattern every serious tool uses
Publishing to Instagram is not a single call. It's two: first you upload a media container (image or video URL, caption, hashtags, location, product tags), then you publish the container. This matters because it gives systems a natural checkpoint. Between "upload container" and "publish container," a good tool can:
- Run a final on-brand score against the rendered asset.
- Confirm the account isn't paused.
- Confirm the scheduled slot hasn't been double-booked.
- Retry on transient failures without duplicating the post.
Approval queue vs autopilot: same pipeline, one flag
People imagine approval mode and autopilot as two different products. They aren't. They're the same publish pipeline with one variable: can a post ship without a human tapping approve? In OnBrand, that variable is autopilot: true|false on the brand row. Everything downstream of it is identical:
- Generate a candidate.
- Score it against the brand kit. If it's below the threshold, retry once with the critique folded into the prompt. Still below? Route to human.
- Above threshold and autopilot is on → schedule directly into the next open slot.
- Above threshold and autopilot is off → land in the approvals queue.
- When the slot arrives, the publisher fires. Metrics start populating. Comments start arriving. The reply layer wakes up.
The reason this matters is that most creators use both modes at once for different content types. Feed images on autopilot, reels in approval, stories on autopilot with a tighter threshold. The pipeline doesn't care.
The guardrails that make autopilot safe
"AI posts to Instagram on its own" is a headline. The reason it doesn't end in disaster is the layer of guardrails underneath it:
Brand-score threshold
Every candidate is scored 0·100 by a vision model against the brand kit · identity match if the render was image-edited from a reference, mood and palette otherwise. Nothing below brands.approval_threshold ever leaves the queue. Set it to 96 and only the top of every batch ships; set it to 80 and the engine ships more often but the ceiling is lower.
Daily cost cap
Autonomous generators without a spend limit are how you accidentally rack up a $400 render bill overnight. Every brand carries a daily_cost_cap_usd. When the sum of today's job costs crosses it, the engine hard-stops and writes a cost_cap_hit event. You get an alert; nothing else happens until tomorrow.
Kill switch
One click flips brands.paused = true. Every worker · generator, scheduler, comment replier · checks this flag before it does anything. Sensitive news cycle, brand crisis, migration window? Pause. Resume when you're ready.
Retry limits
Jobs attempt at most three times. After three consecutive failures the post is marked failed, an event is written, and the scheduler moves on. No infinite loops, no cascading retries.
Simulation mode is a feature, not a workaround
Wiring live Instagram credentials on day one is a bad idea for the same reason wiring live Stripe credentials on day one is a bad idea: you don't know yet what the system is going to do. OnBrand runs the entire loop in simulation when IG_ACCESS_TOKEN is absent. Posts transition through approved → scheduled → posted. Metrics populate with realistic numbers (flagged simulated: true in the metrics JSON). Comments arrive in seven languages. Replies go out. Insights get written. The only difference between simulation and live is one branch inside publishPost. When you're ready, you add the token and flip the switch · everything else you built keeps working.
What NINNA does end-to-end
@ninna.artist is the reference implementation. Every image is generated by the engine from her locked reference. Every caption is written in her voice. Every reel is a Kling image-to-video pass on a rendered still. Every post is scored, and the ones that ship average 94 on-brand. Every comment gets a reply within the warm window, in the commenter's language. She sleeps; the loop doesn't.
Related reading: what is an AI content engine? and how to run a faceless Instagram account on autopilot. Or join the waitlist and put it on your account.