Comparison
Which YouTube transcript MCP server should your agent use?
Arcmira is the search engine for the spoken web. Its MCP server gives an agent six tools over indexed YouTube and podcast transcripts: arcmira_describe returns the arcmira client reference; arcmira_execute_read runs a JavaScript program against that client; arcmira_execute_write adds monitor writes and Premium transcripts; arcmira_youtube_search finds where something was said; arcmira_youtube_transcript reads a whole transcript; arcmira_feedback reports what went wrong. One of its methods returns the full transcript of a single video.
This page puts that server beside the single-video fetchers and the podcast databases that rank for the same search. Their claims are quoted from their own pages with the day we read them. Ours link to the docs page that proves each one.
claude mcp add --transport http --scope user arcmira https://mcp.arcmira.com/mcpOr create a free account key with an email and no card, then send it as a bearer.
curl -X POST "https://api.arcmira.com/v1/signups?src=vs-page" -H "Content-Type: application/json" -d '{"email":"you@example.com"}'From a question to an answer
Two questions a single-video fetcher cannot answer, with the client call an agent makes and the result it gets back. Both results are real, pulled on 2026-09-05, trimmed to the rows shown and not edited.
“Which shows mentioned Ramp in the last 30 days?”
arcmira.momentum("ent_14")
Verdict fading, as of 2026-09-05. 18 mentions in the last 7 days, 82 in the last 30 against 98 in the 30 before.
| Show | Mentions, 30 days |
|---|---|
| TBPN | 15 |
| Invest Like The Best | 11 |
| etn. | 7 |
| John Coogan | 6 |
| This Week in Startups | 5 |
The count covers the shows Arcmira indexes, not the whole internet. The paid-versus-organic split needs a Pro+ plan and the result says so.
“Who sponsors TBPN?”
arcmira.sponsors("UC-DRzaGnL_vtBUpCFH5M0tg")
48 recurring sponsors as of 2026-09-02. The five with the most ad reads:
| Sponsor | Ad reads | Videos | Last read |
|---|---|---|---|
| Ramp | 331 | 296 | 2026-09-02 |
| Figma | 220 | 216 | 2026-09-02 |
| Public | 194 | 181 | 2026-08-10 |
| Vanta | 179 | 175 | 2026-03-31 |
| Linear | 171 | 170 | 2026-03-31 |
A free account key reads this slice. The full list needs a Pro+ plan and the result names it. Docs: commercial intelligence
What you can ask it
Axis by axis
Every Arcmira cell is a number or a fact, with the docs page that proves it. The other two columns restate what the pages above say, read on 2026-09-05.
| Axis | Arcmira | Single-video fetchers | Podcast databases |
|---|---|---|---|
| Transcript quality | The video’s own caption track on any account key, the free one included. On a paid plan, Premium is Arcmira’s own transcript of the audio, with a speaker on every line and speakers named once identified. Docs: The two qualities | The video’s caption track. youtube-transcript.ai returns it as plain text. The local servers return text, timed text, or paragraphs from the same track. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Their own transcripts of podcast audio. Podscan lists get_episode_transcript. Pod Engine transcribes “every English podcast with more than 10 reviews on Apple.” Read 2026-09-05, Podscan, Pod Engine |
| Search across shows | arcmira.search returns spoken slices for one topic across every indexed show in one call, each with a watch link and a publish date. Filters narrow it to passages about an entity or spoken by one person. Docs: Tools | One video per call. None of the six lists a search tool. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Podscan lists search_episodes, “Full-text transcript search.” Pod Engine sells search within its dataset. Read 2026-09-05, Podscan, Pod Engine |
| Sponsors and mentions | arcmira.mentions says whether a show has mentioned a brand and when it first and last did. arcmira.sponsors returns a channel’s recurring sponsors from detected ad reads. arcmira.momentum compares a window you choose, from 7 to 365 days, against the window before it and returns a verdict. Docs: Tools | None. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Podscan lists get_latest_sponsors and alert tools for brand mentions. Pod Engine lists alerts. Read 2026-09-05, Podscan, Pod Engine |
| YouTube and podcasts | Both, in one index with one set of ids. Every client method reads the same catalog. Docs: MCP server | YouTube only. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Podcasts. Pod Engine adds YouTube links and view counts per show, not YouTube transcripts. Read 2026-09-05, Podscan, Pod Engine |
| Auth and price | Sign in through the host over OAuth, or send a key. With an email and no card, two POSTs create a free account key with 1,000 credits a month. The Free plan withholds the newest 30 days of media. Docs: Sign up from the API | youtube-transcript.ai takes no key and no signup. ytmcp.com takes an API key. The local servers take nothing and run on your machine. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Podscan signs you in with a Podscan account over OAuth and meters a daily request quota per plan. Pod Engine sells one plan at $100 a month. Read 2026-09-05, Podscan, Pod Engine |
| Stable ids and dates | Every entity has an ent_ id that does not change. Every result carries as_of, the newest indexed date the answer rests on, so an agent can say what its window was. Docs: Tools | A YouTube video id. No entity ids, no as_of. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Not stated on either page. Read 2026-09-05, Podscan, Pod Engine |
| No deletes | arcmira_describe, arcmira_execute_read, arcmira_youtube_search and arcmira_youtube_transcript declare readOnlyHint true, and every API request their calls send is marked read-only, so none of them starts a Premium transcript. Both execute tools declare openWorldHint true, because a program can name any public video. Only arcmira_execute_write declares destructiveHint true, because it changes monitors and pauses instead of deleting, and it alone starts a Premium transcript. Programs run in an isolate whose only network is the Arcmira API, under a per-tool allowlist. Nothing reads the open web beyond the caption track of a video you name. Docs: Tools | Read-only. Each call fetches YouTube. Read 2026-09-05, youtube-transcript.ai, ytmcp.com, kimtaeyoon83, jkawamoto, sinco-lab, Docker | Podscan lists create_alert, delete_alert, create_list, delete_list, and delete_report among its tools. Read 2026-09-05, Podscan |
| Tools | Six. arcmira_describe returns the arcmira client reference; arcmira_execute_read runs a JavaScript program against that client; arcmira_execute_write adds monitor writes and Premium transcripts; arcmira_youtube_search finds where something was said; arcmira_youtube_transcript reads a whole transcript; arcmira_feedback reports what went wrong. One program can resolve a name and run every query a question needs. Docs: Tools Source on GitHub | One on youtube-transcript.ai and kimtaeyoon83. Four on jkawamoto, sinco-lab, and the Docker image. Read 2026-09-05, youtube-transcript.ai, kimtaeyoon83, jkawamoto, sinco-lab, Docker | 118 on Podscan. Pod Engine does not list a count. Read 2026-09-05, Podscan, Pod Engine |
A host loads every tool description into the agent’s context before its first call. Four tool definitions and the server instructions cost about 2,750 tokens, against 8,163 for the definitions of the ten tools they replaced. The agent reads the client reference once, then writes one program per question.
| The agent asks | Client method |
|---|---|
| Who or what is this name, handle, URL, or channel id? | arcmira.resolve |
| Where was this topic said, across every show, with watch links? | arcmira.search |
| Has this show mentioned this brand, and when first and last? | arcmira.mentions |
| Is this entity accelerating, flat, or fading, over any window from a week to a year? | arcmira.momentum |
| What do these shows talk about, and what do they share? | arcmira.occurrences |
| Who sponsors this channel, by ad-read count? | arcmira.sponsors |
| Who recommends this brand on air, and were they paid? | arcmira.recommendations |
| How much of this channel is indexed, and through what date? | arcmira.status |
| What is the latest episode of this show? | arcmira.episodes |
| What is the full transcript of this one video? | arcmira.transcript |
What each server sells on, quoted 2026-09-05
| Server | Kind | What it sells on | Source |
|---|---|---|---|
| youtube-transcript.ai | Hosted, one video per call | “Free · no API key · no signup · works in any MCP-compatible client.” “No account, no token, no quota wall.” “A hardened fetch pipeline with proxy fallback gets transcripts even on videos that block typical scrapers.” Any caption track by language code. “One clean tool, plain-text output with title and metadata.” Three example prompts. | Read 2026-09-05 |
| ytmcp.com | Hosted, API key | “Download full video transcripts with timestamps for content analysis.” Comments, channel info, trending videos, playlists, and video metadata. “Set up your API key and connect your AI assistant to YouTube.” The page shows no keyless path. | Read 2026-09-05 |
| kimtaeyoon83/mcp-server-youtube-transcript | Local, npm | 592 stars. One tool, get_transcript, taking a URL and a language. “Optional timestamps for referencing specific moments.” “Built-in ad/sponsorship filtering (enabled by default).” “Zero external dependencies for transcript fetching.” MIT. | Read 2026-09-05 |
| jkawamoto/mcp-youtube-transcript | Local, pip or uvx | 477 stars. Four tools, get_transcript, get_timed_transcript, get_video_info, and get_available_languages. Language selection. Splits transcripts over 50,000 characters and hands back a cursor. Proxy support. | Read 2026-09-05 |
| sinco-lab/mcp-youtube-transcript | Local, npm | 36 stars. The same four tools. “Timestamped transcript output.” “Format text with continuous or paragraph mode.” Text normalization and HTML entity decoding. | Read 2026-09-05 |
| Docker MCP catalog, mcp/youtube-transcript | Local, one docker run | “docker run -i --rm mcp/youtube-transcript.” Packages jkawamoto’s server. No API key. MIT, 100K+ pulls. | Read 2026-09-05 |
| podscan.fm | Hosted podcast database | “118 Specialized Tools.” “Over 4,874,431 podcasts and 69,456,623+ episodes.” “Log in with your Podscan account via OAuth.” “Set Up in 2 Minutes.” Setup steps for ChatGPT, Claude.ai, Claude Desktop, and Claude Code above the fold. Customer logos. Daily request quotas from 1,000 to 10,000 by plan. | Read 2026-09-05 |
| podengine.ai | Hosted podcast database | “The world’s first podcast MCP server.” “Millions of podcasts, charts, transcripts, and contact data.” “1 million minutes a day” transcribed. One plan at $100 a month with 1,000 transcripts a month. Use-case videos with prompts you can copy. | Read 2026-09-05 |
The same thirty seconds, two ways
Seconds 45 to 75 of Demonstrating Home Robotics Live on MTS, Moment of Truth, pulled on 2026-09-05. Left is the video’s automatic caption track, the same track every single-video fetcher returns. Right is Arcmira’s Premium transcript of the same audio. Both are shown as returned, trimmed to the window and not edited.
Caption track
- 0:44ago. Um, and I have so many things to
- 0:48ask you about there. But before we get
- 0:50into act two preview, uh we're gonna say
- 0:52would love for you to kind of give the
- 0:54really high level for on Sunday for
- 0:56folks who like aren't familiar. Uh so
- 0:59yeah say thanks for coming on Tony.
- 1:01>> Yeah absolutely. Thank you. Thank you
- 1:03Brexton. Um so at Sunday we're
- 1:05essentially building the full stack
- 1:07intelligence layer robotics and what we
- 1:09mean by that is literally everything
- 1:11like we build our own hardware, we build
- 1:13our own software, we collect our own
Arcmira Premium
- 0:45BraxtonUm, and I have so many things to ask you about there, but before we get into Act 2 preview, uh, we're gonna say would love for you to kind of give the really high level for, uh, on Sunday for folks who aren't familiar.
- 0:59So yeah, I was going to say thanks for coming on, Tony.
- 1:01TonyYeah, absolutely.
- 1:02Thank you.
- 1:02Thank you, Braxton.
- 1:04So at Sunday, we're essentially building the full-stack intelligence layer for robotics.
- 1:09And what we mean by that is literally everything.
- 1:10We build our own hardware, we build our own software, we collect our own data, and of course, we train the model ourselves.
Speaker names come from Arcmira’s speaker identification, which is right most of the time and wrong sometimes. Captions come back on any key. Premium text needs a paid plan, and one Premium read uses credits from the plan, then the on-demand budget, which is the approval. Docs: the two qualities
Install it in your host
The server URL is https://mcp.arcmira.com/mcp. Hosts that speak the MCP authorization spec sign you in, and every other host takes a bearer key. Add it beside the server you use today. Nothing has to be removed to call both on the same video, and arcmira.transcript answers from our cache when we hold the track and fetches YouTube when we do not. Docs: OAuth for MCP clients
ChatGPT
1In ChatGPT on the web, open Plugins, select the plus button, then Add custom MCP server.
2Enter Arcmira as the name. Under Connection, enter the server URL.
https://mcp.arcmira.com/mcp3Choose OAuth for authentication. Review the risk warning, select I understand and want to continue, then select Create as a plugin.
4Find Arcmira in your plugins, install it, and sign in with your Arcmira account.
5In a new chat, type @ and select Arcmira.
Codex
1Add the server from your terminal.
codex mcp add arcmira --url https://mcp.arcmira.com/mcp2Codex opens the Arcmira sign-in in your browser. Complete it, then start a new Codex session. If the sign-in did not open, run codex mcp login arcmira.
3Or install the Arcmira plugin: the same server plus the Arcmira skills. Add the marketplace, then open /plugins in Codex and install Arcmira.
codex plugin marketplace add arcmira/mcpUpdate the plugin with codex plugin marketplace upgrade arcmira, then restart Codex.
Codex CLI, the IDE extension, and the desktop Codex experience share config.toml on the same host. ChatGPT web uses a separate plugin connection.
Claude
1Open Customize, then Connectors, and choose Add custom connector. Some versions place it under +.
2Name it Arcmira and enter the server URL. Leave the headers empty, keep Sign in now and Use Claude’s published identity selected, then choose Add.
https://mcp.arcmira.com/mcp3Complete the Arcmira sign-in. In a conversation, open +, then Connectors, and enable Arcmira.
On Team and Enterprise plans, an owner adds the connector in Organization settings → Connectors. Each member then connects their own account and enables it in the conversation.
Claude Code
1Add the server from your terminal.
claude mcp add --transport http --scope user arcmira https://mcp.arcmira.com/mcp2Run /mcp in the session and follow the browser sign in.
The tools show up in the same session. No restart, no config file to edit by hand.
3Or install the Arcmira plugin: the same server plus the Arcmira skills, such as sponsor research, company watch, quotes and person research.
claude plugin marketplace add arcmira/mcp
claude plugin install arcmira@arcmiraTurn on auto-update, since Arcmira ships weekly and Claude Code leaves it off for third-party marketplaces: run /plugin, open Marketplaces, pick arcmira, and choose Enable auto-update. Update now with claude plugin update arcmira@arcmira.
Cursor
1Install it in one click and confirm the Arcmira server in Cursor.
Add arcmira MCP server to Cursor2Or install it from the Customize page, or write the file yourself. Use .cursor/mcp.json for one project, ~/.cursor/mcp.json for every project, and keep your other servers.
{
"mcpServers": {
"arcmira": {
"url": "https://mcp.arcmira.com/mcp"
}
}
}Use a url entry without a type field.
3Complete Cursor’s OAuth sign-in when prompted, then open Agent and try an Arcmira read.
OpenCode
1Merge this arcmira entry into mcp in your project’s opencode.json or the global ~/.config/opencode/opencode.json. Keep existing servers.
{
"mcp": {
"arcmira": {
"type": "remote",
"url": "https://mcp.arcmira.com/mcp"
}
}
}2Sign in through your browser.
opencode mcp auth arcmira3Open a session and ask OpenCode to make a small Arcmira read.
Hermes Agent
1Add the server with OAuth, in the terminal on the machine running Hermes.
hermes mcp add arcmira --url https://mcp.arcmira.com/mcp --auth oauth2Complete sign-in when Hermes opens your browser. If sign-in stops before it finishes, run the add command again, and answer y when Hermes asks whether to overwrite.
3If Hermes later says the connection needs sign-in, run:
hermes mcp login arcmira4In Hermes, run /reload-mcp or start a new session, then ask for a small Arcmira read.
Optional, on Hermes 0.23.0 or later: Hermes runs a server added this way without asking before any tool. To have it ask before arcmira_execute_write and arcmira_feedback, add trust: untrusted under mcp_servers.arcmira in config.yaml, then run hermes gateway restart --detach and start a new session. Hermes reads that line only when its gateway connects the server, so /reload-mcp or a new session alone keeps the old setting. Search, transcript and program reads still run without a prompt. Scheduled reads run without asking and use credits from your plan. A scheduled run can't answer the prompt, so unless you allowed arcmira_execute_write always, it starts no Premium transcript and changes no monitor.
Grok Bot
1Paste this into your Bot. It adds Arcmira and asks you to sign in.
Add this MCP server: https://mcp.arcmira.com/mcp
It is Arcmira (streamable HTTP, OAuth sign-in), which searches what people said on YouTube. Steps for every host: https://arcmira.com/agent-setup/prompt.md2Approve Arcmira on the sign-in page. If you are new to Arcmira, that page creates a free account.
3Then ask the Bot your question.
Grok
1Open Grok connectors, choose New Connector, then Custom.
2Enter the server URL and complete Arcmira sign-in when prompted.
https://mcp.arcmira.com/mcp3Use the discovered Arcmira tools in a conversation and verify a small read.
On Business and Enterprise, a team admin adds the connector first.
Any other client
Most hosts read an mcpServers object. Give it the URL alone to sign in through the host, or add an Authorization header to send a key.
{
"mcpServers": {
"arcmira": {
"url": "https://mcp.arcmira.com/mcp"
}
}
}An inbox is enough
Create a key from the API
Post an email address and a six digit code lands in that inbox. No browser, no card.
curl -X POST "https://api.arcmira.com/v1/signups?src=vs-page" -H "Content-Type: application/json" -d '{"email":"you@example.com"}'Post the code back and the response carries the key.
curl -X POST "https://api.arcmira.com/v1/signups/verify" \
-H "Content-Type: application/json" \
-d '{"email":"you@example.com","code":"123456"}'The key comes back with an arc_sk_ prefix. It reads the data available to the account. Free accounts get 1,000 credits a month, and the Free plan withholds the newest 30 days of media. Follow the cutoff returned by each tool. Send the key in an Authorization: Bearer header wherever your host lets you set a header. Signing in through the host does the same thing and creates the account on the way.
A free account key is the fastest way to check the index against the fetcher you use today. The same video, and then search, mentions, momentum, and sponsors, which the fetcher does not have. Nothing expires, so the same key is still there the next time you compare.
Questions
Is there a free YouTube transcript MCP server?
More than one. youtube-transcript.ai runs without a key, the local servers run on your machine, and Arcmira creates a free account key from an email address and two POSTs. The difference is what comes back. A fetcher returns a caption track. Arcmira returns an index you can search.
Which MCP server returns speaker labels?
Arcmira’s arcmira.transcript with quality premium, on a paid plan, returns a speaker on every line and names the speakers it has identified. A caption track carries no speakers, so a single-video fetcher cannot add them.
Does Arcmira replace a single-video fetcher?
For the full transcript of one video, yes. arcmira.transcript returns the caption track on any key, from our cache when we hold it and from YouTube when we do not. The other client methods do what a fetcher cannot. They search across shows and report mentions, momentum, sponsors, and coverage.
Give the agent an index, not a fetcher.
Six tools that read indexed YouTube and podcast transcripts and set up monitors. Sign in through your host, or create a free account key.
Read the MCP docs