Arcmira API

YouTube Transcript API

Send a video id, get the transcript back as JSON with a timestamp on every line. Ask for Premium and every line also says who is speaking. Then search what was said across every video and podcast Arcmira has indexed, not only the one you fetched.

A free key reads caption transcripts right away, with no card. The same API powers the Arcmira MCP server, so your agent reads the same data your code does.

Quickstart

Three requests that run as written

Export your key as ARCMIRA_API_KEY and paste any of these. Each runs against production.

Get a YouTube transcript

The video's own caption track as timestamped lines. 4 credits per started 15 minutes of video.

curl
curl "https://api.arcmira.com/v1/transcripts/aHhB3sjGjkI?language=en" \
  -H "Authorization: Bearer $ARCMIRA_API_KEY"

Read who said each line

quality=premium reads a transcript the account owns for 0 credits. Otherwise the GET starts one, using credits from the account's plan, then its on-demand budget, and answers 202 pending with the job. State failed means the last attempt failed, and retry=true starts it again.

TypeScript
const url = "https://api.arcmira.com/v1/transcripts/aHhB3sjGjkI?quality=premium&start=60&end=120"
const res = await fetch(url, {
  headers: { Authorization: `Bearer ${process.env.ARCMIRA_API_KEY}` },
})
const transcript = await res.json()
if (transcript.state === "pending") {
  console.log("pending", transcript.job.status_url, transcript.job.next_poll_seconds)
} else if (transcript.state === "failed") {
  console.log("failed", transcript.last_attempt.error)
} else if (!res.ok) {
  throw new Error(transcript.error?.code ?? `HTTP ${res.status}`)
} else {
  const names = new Map<number, string>()
  for (const s of transcript.speakers ?? []) names.set(s.id, s.name)
  for (const line of transcript.lines) {
    console.log(`[${line.start}s] ${names.get(line.speaker) ?? "Unidentified speaker"}: ${line.text}`)
  }
}

Search what was said across videos

One topic per call. Each result carries the spoken text, a timestamped watch link and the publish date. 4 credits per result past the first 5 of a request.

Python
import json, os, urllib.parse, urllib.request

query = urllib.parse.urlencode({"q": "datacenter power demand", "limit": 3})
request = urllib.request.Request(
    f"https://api.arcmira.com/v1/search?{query}",
    headers={"Authorization": f"Bearer {os.environ['ARCMIRA_API_KEY']}"},
)
with urllib.request.urlopen(request) as response:
    for chunk in json.load(response)["chunks"]:
        print(chunk["cite_line"] or chunk["watch_url"])
        print(chunk["text"][:240], "\n")

What comes back

The Premium response for the TypeScript request, trimmed to two lines. A repeat read of a transcript the account owns uses 0 credits.

Response
{
  "video": {
    "id": "aHhB3sjGjkI",
    "title": "Agents Building Agents - Alfonso Graziano, Nearform",
    "channel_name": "AI Engineer",
    "duration_seconds": 1814,
    "watch_url": "https://www.youtube.com/watch?v=aHhB3sjGjkI"
  },
  "quality": "premium",
  "source": "arcmira_premium",
  "lines": [
    { "start": 60.54, "end": 67.42, "speaker": 0, "index": 14,
      "text": "So today we will see how we are building and improving iteratively AI agents." },
    { "start": 67.84, "end": 71.77, "speaker": 0, "index": 15,
      "text": "And of course, as you may guess, how can we do that, right?" }
  ],
  "speakers": [{ "id": 0, "name": "Speaker 1", "entity_id": null, "confidence": null }],
  "revision": "r1hk7or3.jx73u4",
  "rows_billed": 0
}

Two lanes

Captions for everyone, Premium when it matters

Captions

Any key, including a free one

4 credits per started 15 minutes of video

  • The video's own caption track: creator captions when the channel wrote them, YouTube's automatic captions otherwise. The source field says which.
  • Timestamped lines by default. timestamps=false returns paragraphs for reading instead of citing.
  • start and end return one window of the video and charge only for that window.

Premium

Paid plans and the Pro trial

300 credits per started 15 minutes, then free for that video forever

  • Arcmira's own transcript, made from the audio and diarized. Every line carries a speaker id.
  • speakers[] names each voice once someone has identified it, with a confidence, and says Speaker 1, Speaker 2 when nobody has.
  • GET /v1/transcripts/{id}?quality=premium returns the transcript the account owns, a pending job (202), or a failed state. When the account does not own it yet, the read starts it for the whole video, using credits from the account's plan, then its on-demand budget. After a failure, retry=true starts it again.

Pick the lane per request with quality=captions or quality=premium, or set an account default once with PATCH /v1/me/settings. The transcripts guide has every parameter.

Beyond one video

More than a caption fetcher

Search what was said

GET /v1/search

Find spoken passages across indexed YouTube and podcast transcripts by topic, then scope by channel, person or date. Each result carries a timestamped watch link and a ready citation line.

Follow who was mentioned

GET /v1/mentions

Every indexed moment a company, person, product or topic came up, with the video, the second it starts, and the sentiment. Filters take entity ids: resolve a name with GET /v1/entities/resolve first.

Give it to an agent

mcp.arcmira.com/mcp

The same transcripts, search and mentions as an MCP server for Claude, ChatGPT, Cursor and Codex, with OAuth sign-in through the host.

Connect an agent in one step from the setup guide, or read what the MCP server does.

Languages

Every caption track the video has

language is a priority list of up to 5 codes, tried in order, such as de,en. Use asr for the first automatic track or asr-de for a specific one. The response names the track it resolved and lists every track on offer, so a second call can ask for exactly the right one.

The example video above lists 21 tracks, from Arabic and Bangla to Tamil and Ukrainian. Coverage is per video: it is whatever the channel and YouTube provide.

Pricing

Pay in credits, not requests

Reads use credits from your plan, the same balance the web app draws from. A free account gets 1,000 credits a month and needs no card. Paid plans start at $20 a month and refresh their credits every month; past them, on-demand usage runs up to a budget you set.

RequestCost
Caption transcript4 per started 15 minutes, minimum 4
Premium transcript300 per started 15 minutes, once per video
Transcript search4 per result past the first 5 of a request
/v1/me, health, OpenAPI0
The same request again within 7 days0
A 404 or a 5030

Compare plans or read how credits work. GET /v1/me reports what is left, for free.

Compared

What the caption track gives you, and what Arcmira adds

NeedCaption track aloneArcmira
Timestamped linesYesYes, on both lanes
Pick a languageTracks the video offersA priority list of up to 5 codes, automatic tracks with asr
Who said each lineNoA speaker id on every Premium line
Speaker namesNoNamed once identified, with a confidence
Videos without captionsNothing to returnPremium transcribes the audio
Search by what was saidOne video at a timeAcross every indexed video and podcast
Who was mentioned whereNoEntity mentions with timestamps
Fix a wrong lineNoCorrections anchored to a revision and line index

Choosing an MCP server instead? See the YouTube transcript MCP server comparison.

Built for production

Things you will not have to guess

One contract
An OpenAPI 3.1 document at https://api.arcmira.com/v1/openapi.json describes every route, parameter and error. The version is /v1 in the path.
Readable limits
Every response carries RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset. A 429 carries Retry-After. Free keys get 60 requests a minute, paid plans 240 or more.
Typed errors
Every error has a stable code and a doc link, and a plan gate names the plan that lifts it. The error catalog lists them all.
Safe retries
Every response has an X-Request-Id. Every write accepts an Idempotency-Key, so a retried request returns the stored response instead of acting twice. Repeated Premium reads join the same job and never charge twice.
Open source agent surface
The MCP server is public on GitHub and listed in the MCP registry as io.github.arcmira/arcmira.

FAQ

Questions developers ask

Is there a free YouTube transcript API?

Yes. A free Arcmira account gets 1,000 credits a month and needs no card. A caption transcript uses 4 credits per started 15 minutes, so an hour-long video is 16 credits. Create the key in the browser or with two API calls.

What does a transcript response contain?

The video (id, title, channel, publish date, duration, watch link), the lane that answered, the source class, the resolved language, the caption tracks on offer, and lines of start, end and text. Premium lines add a speaker id and a line index, and the response adds speakers and a revision.

Which languages does it support?

Whatever caption tracks the video offers. Send language as a priority list of up to 5 codes, such as de,en. Use asr for the first automatic track or asr-de for a specific one. Every transcript response lists the tracks on offer.

Can I get speaker labels?

Yes, on Premium transcripts. On a paid plan or the Pro trial, a Premium transcript of the whole video uses credits from your plan, then your on-demand budget. Every line carries a speaker id, and speakers lists a name for each voice once it has been identified. Identification is right most of the time and wrong sometimes, so cite the line when a name matters.

What happens when a video has no captions?

The captions lane answers transcript_unavailable. On a paid plan or the Pro trial, GET /v1/transcripts/{video_id}?quality=premium makes a Premium transcript from the audio and answers pending until it is ready. A failed transcript gives its credits back, and retry=true starts a new one.

Does it cover podcasts?

Transcript search covers indexed YouTube and podcast conversations. GET /v1/transcripts/{video_id} takes a YouTube video id.

Can an AI agent call it?

Yes. Point any MCP client at https://mcp.arcmira.com/mcp and sign in through the host, or send an account key as a bearer header from a script.

Your first transcript is one request away.

Sign up, copy the key, run the curl above. Everything else is in the developer portal.

Get a free API key