Read the text shown on screen in a video
Read the words written on screen in an Instagram reel or a TikTok, second by second, and the voiceover beside them
Send a reel or TikTok link and get back the words shown on screen, with the second each one appears. Add the transcript when you also need what is said. The examples use two public NASA videos.
Measured on 18/11/2026: the 29-second reel showed 8 lines of text, from "THIS OR THAT" at 0 seconds to an application link at 25. Its voiceover was one short exclamation. The three steps cost 30 credits.
Start in five minutes
Set your key once. Every block on this page reads it.
export SOCIALCRAWL_API_KEY=YOUR_API_KEYPaste this prompt into your coding agent:
Read https://www.socialcrawl.dev/docs/recipes/on-screen-text.md and follow it to read the on-screen text of https://www.instagram.com/reel/DfpowZdlfMz/. Use the key in SOCIALCRAWL_API_KEY. Price each paid call with a dry run first (add dry_run=1, which is free), and add up the prices. Keep the whole job under max_credits=40: stop and ask me before you go past it.Or price the first call yourself. A dry run validates the request for 0 credits:
curl -G "https://www.socialcrawl.dev/v1/instagram/media/screen-text" \
-H "x-api-key: $SOCIALCRAWL_API_KEY" \
--data-urlencode "url=https://www.instagram.com/reel/DfpowZdlfMz/" \
--data-urlencode "frames=all" \
--data-urlencode "dry_run=1"It answers a hold of 10 credits, settling between 5 and 10.
How it works
- Captions and transcripts miss the words a creator writes on the video itself.
- The screen-text call reads those words off the video's frames. On TikTok it also reads the app's own text stickers.
- By default it reads the cover. With
frames=allit reads the whole video, up to three minutes. - The transcript call returns what is said, so you can tell a spoken name from a written one.
- You read each post once and keep the answer.
What you need
| Call | What it gives | Credits |
|---|---|---|
GET /v1/instagram/media/screen-text | Text on a reel's frames, or on a carousel's slides | 5, or 10 with frames=all |
GET /v1/instagram/media/transcript | What is said in a reel | 10 |
GET /v1/tiktok/video/screen-text | Text on a TikTok video's frames, or on a photo post's slides | 5, or 10 with frames=all |
GET /v1/tiktok/post/transcript | What is said in a TikTok | 10 |
Steps
Step 1: the text on a whole reel
curl -G "https://www.socialcrawl.dev/v1/instagram/media/screen-text" \
-H "x-api-key: $SOCIALCRAWL_API_KEY" \
--data-urlencode "url=https://www.instagram.com/reel/DfpowZdlfMz/" \
--data-urlencode "frames=all"import json, os, urllib.parse, urllib.request
params = {"url": "https://www.instagram.com/reel/DfpowZdlfMz/", "frames": "all"}
url = "https://www.socialcrawl.dev/v1/instagram/media/screen-text?" + urllib.parse.urlencode(params)
req = urllib.request.Request(url, headers={"x-api-key": os.environ["SOCIALCRAWL_API_KEY"]})
data = json.load(urllib.request.urlopen(req, timeout=180))["data"]
print(data["texts"])
for block in data.get("timeline") or []:
print(block["at_seconds"], block["kind"], block["text"])textsholds every distinct line once, in reading order.timelinegives each line withat_seconds, the second it first appears, andkind.subtitlesholds the on-screen captions of the speech, kept apart fromtexts.
{
"texts": ["THIS OR THAT\nASTRONAUT CANDIDATES", "Earthrise 🌍 Moonrise 🌕"],
"timeline": [
{ "at_seconds": 0, "kind": "text", "text": "THIS OR THAT\nASTRONAUT CANDIDATES" },
{ "at_seconds": 10, "kind": "text", "text": "Earthrise 🌍 Moonrise 🌕" },
{ "at_seconds": 25, "kind": "text", "text": "Apply to be\nan astronaut at:\nGO.NASA.GOV/ASTRO2026" }
],
"subtitles": [],
"duration_seconds": 29.117,
"frame": "all"
}Step 2: what is said in it
curl -G "https://www.socialcrawl.dev/v1/instagram/media/transcript" \
-H "x-api-key: $SOCIALCRAWL_API_KEY" \
--data-urlencode "url=https://www.instagram.com/reel/DfpowZdlfMz/"import json, os, urllib.parse, urllib.request
params = {"url": "https://www.instagram.com/reel/DfpowZdlfMz/"}
url = "https://www.socialcrawl.dev/v1/instagram/media/transcript?" + urllib.parse.urlencode(params)
req = urllib.request.Request(url, headers={"x-api-key": os.environ["SOCIALCRAWL_API_KEY"]})
data = json.load(urllib.request.urlopen(req, timeout=180))["data"]
for t in data["transcripts"]:
print(t["text"])- A transcript is kept for 30 days, so asking again in that time is a free cache hit.
- On TikTok,
GET /v1/tiktok/post/transcriptdoes the same job for 10 credits.
Step 3: the same on TikTok
curl -G "https://www.socialcrawl.dev/v1/tiktok/video/screen-text" \
-H "x-api-key: $SOCIALCRAWL_API_KEY" \
--data-urlencode "url=https://www.tiktok.com/@nasa/video/7705995616794561805" \
--data-urlencode "frames=all"import json, os, urllib.parse, urllib.request
params = {"url": "https://www.tiktok.com/@nasa/video/7705995616794561805", "frames": "all"}
url = "https://www.socialcrawl.dev/v1/tiktok/video/screen-text?" + urllib.parse.urlencode(params)
req = urllib.request.Request(url, headers={"x-api-key": os.environ["SOCIALCRAWL_API_KEY"]})
data = json.load(urllib.request.urlopen(req, timeout=180))["data"]
print(data["duration_seconds"], "seconds")
for block in data.get("timeline") or []:
print(block["at_seconds"], block["kind"], block["text"])native_textsholds the text from TikTok's own text tool, andocr_textswhat was read off the frames.sourcesays which of the two produced the answer.
What you get back
The two examples came back differently. One showed only written text, and the other only captions of its speech. On the reel, every line on screen was written by the creator: 8 lines in texts, and no subtitles. The voiceover said none of them. On the 61-second TikTok, every line on screen was a caption of the speech. So texts came back empty and subtitles held 48 lines, such as "THE MEATBALL BETTER" at 31 seconds.
Store only what is new
Key each answer on the post's URL and never read a post twice. When you find posts with a search, store the URLs first, and send the screen-text call only for URLs you have not read. Keep texts and timeline as they came, and run your own matching (place names, prices, product names) over them later, for free.
What it costs
| Read | Credits |
|---|---|
| A reel or TikTok cover | 5 |
A whole reel or TikTok, up to three minutes, with frames=all | 10 |
| Ten slides of a carousel or photo post | 5 |
| A transcript | 10 |
| A dry run | 0 |
What the example runs measured on 18/11/2026:
| Run | Calls | Credits |
|---|---|---|
| The three steps, first read | 3 | 30 |
| The full program straight after, two videos with both transcripts | 4 | 10 |
The second run paid only for the TikTok transcript. The other three reads were cache hits, which are free.
Tips, and what not to use
- A carousel is read 10 slides a call. Send
slide_offset=10for slides 11 to 20, and keep going whilenext_slide_offsetis set. - With
frames=all, a video longer than three minutes is refused and refunded. Read its cover instead, withoutframes. - When a video cannot be read whole, the cover is read, the call costs 5, and
warningssays so. - A post with no text on screen answers an empty
textslist. That is a real answer, not an error. - Do not use a transcript to find a place name shown on screen. Speech and screen can carry different words.
Full program and agent brief
Next steps
Site links
Search X, Reddit and the web for posts that link to one domain, read each thread once, and keep every link the first time it appears
Campaign creators
Search TikTok, Instagram and YouTube for creators in one market above a follower floor, priced first with a free dry run, and keep each one once
