ALL WORK

CASE STUDY · OWN PRODUCT · MCP SERVER

Every meeting you ever recorded, one question away

Soniox Smart Scribe records my meetings and transcribes them almost perfectly — then keeps the answers locked in an app with no public API. So I reverse-engineered the private backend the web app itself uses and built an MCP server: Claude searches, reads and quotes my meetings with speaker labels and timestamps, locally or from claude.ai and my phone. It is open source on GitHub.

ROLE
Reverse-engineering · build · packaging
PRODUCT
MCP server for Claude
STACK
TypeScript · MCP SDK · Vercel
STATUS
Open source · live

YOU What did we agree with the supplier last week?

⚙ soniox_list_transcripts · last 7 days → 4 meetings

⚙ soniox_search_in_transcript · “supplier” → 6 hits

CLAUDE Three commitments from Tuesday’s call:

— delivery moves to 15 January [12:40]

— they cover shipping above €500 [13:05]

— you send revised specs by Friday [21:18]

The whole product in one exchange: a question in plain language, two tool calls, an answer with timestamps you can check against the recording.

01 / THE PROBLEM

Smart Scribe remembers everything and tells you nothing

I record meetings with the Soniox app because its transcription is the best I have used — speaker labels, Slovene and English in the same sentence, timestamps to the second. But the question I actually have a week later is never about one meeting. It is what did we agree, across four calls, two of which I barely remember. The app answers with a list of recordings and a search box that looks inside one transcript at a time.

The obvious fix is to let an AI assistant read the transcripts. Claude already does this well for anything it can reach over MCP — but Soniox has a public API only for its speech-to-text engine, not for the app where my meetings live. Officially, the data is reachable by me, in a browser, one transcript at a time, and by nobody else.

02 / THE API

The API existed. It just wasn’t published.

The web app at app.soniox.com has to get the transcripts from somewhere. DevTools showed where: a private backend, mobile-app-backend.soniox.com, speaking clean JSON. I wrote a pair of throwaway Python probes — they are still in the repo — and mapped the shapes: how transcripts list, how paragraphs and speaker turns nest, where the token lives. The login is standard OAuth2 with PKCE, and the browser keeps {accessToken, refreshToken, expiresAt} in localStorage, one copy-paste away.

That made the plan simple: a TypeScript MCP server that signs in the same way the web app does and exposes the transcripts as tools. Local first, as a stdio server for Claude Desktop and Claude Code; then the same core behind a stateless HTTP endpoint on Vercel, so claude.ai and my phone could use it as a custom connector.

03 / THE HARD PART

A token that self-destructs every time you use it

The access token lives one hour. The refresh token rotates on every use, and the old one is revoked minutes after rotation. That is good security design and a trap for exactly what I was building: a serverless function that refreshes a token and forgets the new one has just locked itself out — and logged you out of the app with it.

The rule that fixed it: whoever refreshes must persist. The server picks a token store from its environment — a private Vercel Blob when deployed (so the serverless endpoint and a local server can share one chain), a JSON file for local-only use, memory as a last resort. Seeding a new login re-seeds the whole chain. It is the least visible part of the project and the one everything depends on.

Two things broke along the way, neither where I was looking:

04 / THE PRODUCT

Six tools, read-only by design

The server deliberately cannot delete, rename or share anything. Four read tools cover the questions; two opt-in tools let the local server upload a recording and transcribe it through your own Soniox quota:

soniox_list_transcripts

List and full-text search transcripts, with date, project and tag filters

soniox_get_transcript

Full transcript as speaker-labelled paragraphs with [mm:ss] timestamps

soniox_search_in_transcript

Find a phrase inside one meeting; timestamped hits with context

soniox_get_workspace_info

Who is signed in, projects and tags — the ids everything filters by

soniox_transcribe_file

Opt-in: upload an audio or video file and transcribe it into Soniox

soniox_get_transcription_status

Opt-in: progress of a running file transcription


In practice, the tools disappear and what remains is asking:

05 / THE PACKAGE

The code is free. The afternoon is the product.

I open-sourced the server and posted the build on LinkedIn; people asked for setup help in the comments, and I found myself writing the same careful DM — token, Node, config paths, the rotation trap — again and again. That DM is the real product. Not the code, which stays free, but the packaged version of the afternoon it saves: a skill your agent runs that sets everything up for you, plus updates when the undocumented backend shifts. That version is in the works.

DO IT YOURSELF

Free

The full server, MIT-style open. Clone it, build it, follow the README. Everything on this page works — you do the setup, and the token trap is yours to discover.

VIEW ON GITHUB

PACKAGED SKILL · IN THE WORKS

Soon

  • The skill packaged for your agent: paste one thing, answer one question, and it walks you through token, install and config on your machine
  • Both setups covered: local for Claude Desktop and Claude Code, remote on your own free Vercel account for claude.ai and your phone
  • The token-rotation trap explained before it bites: one login, one persisted chain, no mystery logouts
  • Updates when Soniox changes the backend — the part a ZIP from GitHub can’t promise
GET NOTIFIED

Your Soniox account, your machine or your Vercel — no service of mine in the middle.

06 / HONEST NOTES

Built on an API that owes me nothing

This talks to an undocumented backend. Soniox can change it tomorrow and something here will break; when it does, I fix the server because I use it every day — that is the only guarantee on offer, stated plainly. The token you copy grants full access to your transcripts, so the README treats it like a password, and the server ships read-only so the worst a misbehaving prompt can do is read.

If Soniox ships a real public API for the app, this project gets smaller and better the same week. I would consider that a win.

07 / QUESTIONS

What people ask before the first question to Claude

What is the Soniox MCP server?

An MCP (Model Context Protocol) server that connects AI assistants like Claude to your Soniox App / Smart Scribe meeting transcripts. Claude can list, search and quote your meetings with speaker labels and timestamps, and optionally transcribe new audio or video files through your Soniox account.

Does it work with claude.ai, Claude Desktop and on my phone?

Yes. Locally it runs as a stdio MCP server for Claude Desktop and Claude Code. Deployed to a free Vercel account it becomes a custom connector for claude.ai, which also makes it available in the Claude mobile apps. Any MCP-capable client can use it.

Is this an official Soniox product?

No. Soniox has no public API for the app, so this talks to the same private backend the web app uses. It is an independent, unofficial tool: it can break if Soniox changes their backend, and it is not endorsed or supported by Soniox.

Is it safe? Where do my transcripts go?

The server is read-only by default and runs on your machine or your own Vercel account — there is no third-party service in between, and nothing is stored outside your Soniox account. The auth token grants full access to your transcripts, so you keep it private, exactly like a password.

The code is free on GitHub. What will the packaged version add?

A guided setup your agent runs for you (token, install, local or remote config), the token-rotation pitfalls pre-solved, and updates when the undocumented backend changes. It packages the saved afternoon, not the code — the code stays free. It is in the works; email me to hear when it ships.

Can it transcribe new recordings, not just read old ones?

Yes, as an opt-in on the local server: point it at an audio or video file and it uploads to Soniox and transcribes with your account’s quota, with ffmpeg audio extraction for large videos. The result appears in the Soniox app like any other transcript.

MY ROLE

API REVERSE-ENGINEERINGTYPESCRIPT MCP SERVEROAUTH2 TOKEN CHAINVERCEL SERVERLESSPRODUCT PACKAGING
BACK TO ALL WORK