Hosted YouTube transcript API, or run your own

What a hosted YouTube transcript API buys you

Reading captions from YouTube is not hard; reading them from a server, every day, is. The library does the parsing. Everything else is operations, and that is what a hosted service is selling.

• Residential exits. YouTube blocks caption requests from cloud IP ranges. We fetch through a residential pool pinned to one region, so RequestBlocked and IpBlocked stay on our side of the line.
• Format changes. YouTube changes the player response shape a few times a year. When it does, we fix the parser; you do not redeploy.
• Caching. Transcripts are cached indefinitely. A second request for the same video, from any of your services, costs nothing. Search results are cached for an hour, channel and playlist listings for six.
• Refunds and structured errors. A video with captions off, private, deleted or age-restricted returns a 404 with retryable: false and a hint, and the credit comes back. A block or an upstream failure is a 5xx, also refunded.
• Uptime. The proxy vendor going down, the pool getting burned, the consent wall in the EU: those are our pages, not yours. GET /health is free and needs no key.
• The rest of YouTube. Search, channel uploads, channel search and playlists sit next to /transcript, one credit each, and a free webhook for new uploads.

What it does not buy: video or audio download, comments, rich metadata, or anything that needs the channel owner's login. And captions must exist. There is no speech-to-text fallback.

The real cost of self-hosting

The line item people forget is not the compute. It is the proxy bandwidth and the hours.

A transcript fetch through youtube-transcript-api pulls the whole watch page, not just the caption track. On the wire that is roughly 300 to 400 KB per successful fetch; decompressed the page is closer to 1.5 MB. Residential proxies are metered on bandwidth, and a blocked attempt pays for the page again. The library's default of ten retries can turn one bad video into 10 to 15 MB. Budget at least 0.5 MB per transcript once retries and consent-wall redirects are counted.

Residential bandwidth pricing moves; as of September 2026 the mainstream vendors quote somewhere in the low single-digit dollars per gigabyte on small plans, falling with commitment. The numbers below use a conservative $3 to $8 per GB range so you can plug in your own quote.

Then the hours. Someone notices the error rate climb, works out whether it is a burned pool or a format change, bumps the library, redeploys, and re-runs the failed batch. Call it two to four hours in a quiet month and a day when YouTube ships a change. At a contractor rate that is the largest number on the self-hosted side.

Cost at 1,000, 25,000 and 200,000 transcripts a month

Self-hosted assumes 0.5 MB of proxy traffic per transcript including retries, $3 to $8 per GB, and a small server. Hosted uses our list prices: 1 credit per transcript, cache hits free. Engineering time is listed separately because your rate is not ours to guess.

The bandwidth column alone is close to the hosted price at every row, before the hours. Where self-hosting wins is at very low volume from a residential connection, where there is no proxy at all, or when video IDs cannot leave your network. Where hosted wins is everywhere a proxy vendor is on the invoice. Repeated videos tilt it further: a summarizer that gets the same trending video from many users pays us once.

Current plans, top-ups and the translation surcharge (+1 credit per 40 segments, minimum 2) are on the pricing page. Other paid services are lined up in the YouTube transcript API comparison.

What you give up

• Control over the fetch path. You do not choose the exit country, cannot supply cookies from your own account to read age-restricted videos, and cannot tune retries. If a video is unreadable for us, it is unreadable, and the hint will say so.
• Per-call pricing. A credit is spent per transcript, search or listing. Failures are refunded and cache hits are free, but a bill that scales with calls is a different shape from a bill that scales with gigabytes.
• A third party in the path. Video IDs and the transcript text pass through us. If that is not allowed where you work, self-host.
• Narrow scope. No downloads, no comments, no descriptions or tags, no analytics, no writes. If you need those, keep the Data API alongside.

If the library is what you are replacing, the youtube-transcript-api alternative page has the side-by-side code and the honest comparison table.

Three ways in: REST, MCP, n8n

The REST API is one bearer token and a GET. Every other route is a wrapper around it.

curl "https://api.transcriptapi.io/transcript?video_id=dQw4w9WgXcQ&format=srt" \
  -H "Authorization: Bearer ta_..."

# format: json (default) | txt | srt | vtt
# language=en, translate_to=de also accepted

• REST. /transcript, /search, /channel/videos, /channel/search, /playlist, and POST /channel/subscribe for new-upload webhooks. Full parameters in the API reference.
• MCP. An MIT-licensed server, npx -y transcriptapi-mcp locally or hosted at api.transcriptapi.io/mcp, gives claude.ai, ChatGPT, Claude Code, Cursor and VS Code five tools: transcript, search, channel videos, channel search, playlist. Setup per client is on the MCP page.
• n8n. An HTTP Request node with the bearer header is all it takes; the n8n guide has a ready-made workflow for summarizing new uploads from a channel.

Frequently asked questions