Gladia MCP removes that detour. It’s an open-source MCP for speech-to-text that connects Claude, ChatGPT and Cursor to your Gladia account, so the assistant can send audio files for transcription and keep the result in context.
TL;DR
- Gladia MCP is an open-source MCP server (MIT license) that runs with one npx command in Claude Desktop, Claude Code, Cursor and the ChatGPT desktop app.
- One prompt can request a transcript with speaker labels plus a summary, translation, SRT or VTT subtitles or PII redaction, processed by Gladia.
- The assistant can also check and delete pre-recorded jobs and live sessions, so monitoring happens in the chat instead of a dashboard.
- Jobs bill to your Gladia account: as of October 2026, async starts at 0.16€ per hour, with 50€ in free credits for new accounts.
How the Model Context Protocol connects assistants to external tools
The Model Context Protocol (MCP) is an open standard that Anthropic introduced in November 2024 to connect AI assistants to external services and databases. An MCP host, such as Claude Desktop or Cursor, runs one or more MCP servers. Each MCP server describes its tools, and the model decides when to call them. We cover the protocol in more depth in what MCP is and why it matters for audio.
What a speech MCP server adds
A speech MCP server exposes transcriptions as tools the model can call. Instead of you running a command line script on your computer, the assistant picks the right tool and fills in the parameters from your prompt. Gladia MCP exposes eight MCP tools:
| Tool |
What it does |
upload_audio |
Uploads a local file from a trusted folder |
transcribe_pre_recorded |
Starts an async, billable transcription job and returns its ID |
get_pre_recorded_transcription |
Reads a job’s status and result |
list_pre_recorded_transcriptions |
Lists jobs with filters and pagination |
delete_pre_recorded_transcription |
Permanently deletes a job and its data |
get_live_transcription |
Reads an existing live session’s status and result |
list_live_transcriptions |
Lists existing live sessions |
delete_live_transcription |
Permanently deletes a live session and its data |
What Gladia MCP can do from a single prompt
Everything below runs on Gladia’s speech-to-text API, with your request passed as options on one pre-recorded (async) job.
Transcribe a recording with speakers and timestamps
Ask for a transcript and the assistant starts an async transcription job, then checks it until it’s done. Add “with speaker diarization” and the transcript comes back split by speaker. Speaker diarization labels who said what: without it, you can’t tell the customer’s objection from the sales rep’s answer. On the DIHARD III benchmark, Solaria-1 has the lowest diarization error rate of the 10 providers we tested, at 16.6% weighted average, according to our open benchmarks.
Summarize, translate and subtitle in the same request
Audio intelligence features ride along on the same job. A prompt like “transcribe this interview, summarize it as bullet points and translate it into French” turns on summarization and translation as options. Ask for both subtitle formats and you get two files, SRT and VTT, ready to drop next to a video.
Redact sensitive data before it reaches your context
Call recordings often contain card numbers and addresses that shouldn’t sit in an assistant’s context. Turning on PII redaction masks that data before the assistant reads it. Sentiment analysis, named entity recognition and custom vocabulary for product names and jargon are available the same way.
Monitor transcription jobs and live sessions without a dashboard
“Which of my transcription jobs failed today?” is now a chat question. The assistant calls the list and get tools and reports each job’s status. Live sessions started from the SDK or your own app show up too, so a developer can check a real-time session’s result without leaving the editor.
Ask across every file you’ve transcribed
Finished transcripts stay reachable, so you can ask “across the last ten customer calls, what’s the most common reason people cancel?” and the assistant pulls each transcript to find the pattern.
How to install Gladia MCP in Claude, ChatGPT and Cursor
You need Node.js 22.18 or later, a Gladia API key from app.gladia.io and an MCP-compatible host. The MCP server runs through npx, so there’s nothing to implement and no global install. Store the API key as an environment variable in the host’s configuration file and never commit it, since anyone holding it can access your account. Then follow the instructions for your host.
Cursor
Add this to .cursor/mcp.json in your project, or to Cursor’s MCP settings, then restart Cursor:
{
"mcpServers": {
"gladia": {
"command": "npx",
"args": ["-y", "@gladiaio/mcp"],
"env": {
"GLADIA_API_KEY": "your-api-key"
}
}
}
}
Claude Code and Claude Desktop
Claude Code takes one command:
claude mcp add gladia --env GLADIA_API_KEY=your-api-key -- npx -y @gladiaio/mcp
For Claude Desktop, use the same JSON shape as the Cursor example in Claude Desktop’s MCP config, then restart the app.
ChatGPT
Gladia MCP runs in the ChatGPT desktop app, which stores MCP servers in the same config file as the Codex CLI and IDE extension. Add this to ~/.codex/config.toml:
[mcp_servers.gladia]
command = "npx"
args = ["-y", "@gladiaio/mcp"]
[mcp_servers.gladia.env]
GLADIA_API_KEY = "your-api-key"
Then open Settings, select MCP servers, and select Restart. Type /mcp in the composer to check that Gladia is connected. If you have the Codex CLI installed, this command writes the same configuration:
codex mcp add gladia --env GLADIA_API_KEY=your-api-key -- npx -y @gladiaio/mcp
ChatGPT on the web doesn’t read local MCP configuration, so use the desktop app.
Local audio files and trusted folders
Public HTTPS links go straight to Gladia. For local files, upload_audio reads from trusted folders only, your Desktop by default. Set GLADIA_LOCAL_FILE_ROOTS to choose other folders (separated by a colon on Mac and Linux, a semicolon on Windows), or to an empty value to turn local uploads off. Files must be under 1 GB, and symlinks are resolved before the folder check so a link can’t escape the allowed root.
What happens between your prompt and the transcript
A transcription request is asynchronous. When you type “Transcribe https://example.com/interview.mp3 with speaker diarization and a bullet-point summary”, the flow looks like this:
- The assistant calls transcribe_pre_recorded with your options. Gladia accepts the job and returns its ID right away.
- The assistant calls get_pre_recorded_transcription until the status is done or error.
- The finished response holds the transcript and every add-on you asked for, as structured JSON. If the job fails, the error message the assistant sees is sanitized, with no credentials or transcript payloads in it.
Note that each transcription job is billable to your account.
Hosted Gladia MCP vs a local faster-whisper MCP server
A common way to build a speech MCP server is to run Whisper locally, often through faster-whisper, a reimplementation of OpenAI’s Whisper that its maintainers report is up to 4 times faster for the same accuracy.
|
Local faster-whisper MCP |
Gladia MCP |
| Where audio processing runs |
Your machine |
Gladia’s cloud (separate EU and US infrastructure) |
| Works offline |
Yes |
No |
| Per-hour cost |
None, beyond your hardware |
From 0.16€ per hour (October 2026) |
| Speaker diarization |
Needs a separate model |
Built in |
| Summaries, translation, PII redaction |
Build them yourself |
Options on the same job |
| Job history and monitoring |
None by default |
List, get and delete tools |
| Setup |
Model download, GPU or CPU tuning |
One command and an API key |
Pick a local server if recordings can’t leave the machine or you need offline transcription. Pick Gladia MCP if you want diarization, audio intelligence and job history without running models yourself. On real customer English audio, Solaria-3 has the lowest word error rate (WER, the share of words a model gets wrong) in our benchmark at 9.6%, ahead of ElevenLabs Scribe v2 at 9.9% and AssemblyAI at 10.0%, all tested on the same files.
Current limits of Gladia MCP
This first release (v0.1.1) is deliberately small, with a few gaps to plan around.
- It doesn’t start or stream live sessions. If you’re building a voice assistant that handles voice interaction while the user talks, the real-time API is still the integration to use.
- It doesn’t expose Audio-to-LLM, custom spelling or callbacks, and the transcription tool has no option to choose between Solaria-1 and Solaria-3.
- It runs over stdio only, with one Gladia account per process. There’s no hosted HTTP or OAuth transport, which is why it works in the ChatGPT desktop app but not in ChatGPT on the web.
- Pre-recorded files are capped at 135 minutes and 1,000 MB, per the supported formats page, so longer recordings need splitting.
- Paid accounts run 25 jobs in parallel and free accounts 3, per the concurrency docs.
It’s also speech-to-text only. The MCP server doesn’t capture voice input from a microphone, and speech generation (text to speech or TTS) isn’t part of it.
Data, security and cost stay on your Gladia account
The MCP server is a thin layer over the official Gladia SDK, so the same data handling applies as for any API call. Gladia holds compliance across GDPR, HIPAA, SOC 2 Type 2, and ISO 27001, with EU and US hosting options on separate infrastructure. Self-serve accounts can opt out of model training for free; Enterprise is opted out by default and offers zero data retention.
Every transcription job bills to the account behind your API key. The following are the Self-serve rates on the pricing page:
| Item |
Rate |
| Async transcription, Solaria-3 |
0.16€ per hour |
| Async transcription, Solaria-1 |
0.27€ per hour |
| Summarization add-on |
+0.03€ per hour |
| Translation add-on |
+0.05€ per hour |
| PII redaction add-on |
+0.07€ per hour |
Async rates drop automatically as monthly volume grows. Speaker diarization and custom vocabulary are included at no extra cost.
Audio is now one more source your agent can read
Gladia MCP doesn’t change what transcription is. It changes where you ask for it: recordings that used to wait in a folder become something Claude, ChatGPT or Cursor can read and compare in the middle of the work.
To try it, create a Gladia account at app.gladia.io to get your API key and 50€ in free credits, roughly 312 hours of async transcription on Solaria-3, then run the install command for your host from the Gladia MCP repository.
FAQs
What is Gladia MCP?
Gladia MCP is an open-source Model Context Protocol server, published as @gladiaio/mcp, that connects AI assistants such as Claude, ChatGPT and Cursor to a Gladia account. It lets the assistant start pre-recorded transcription jobs and check, list or delete jobs and live sessions from a chat prompt.
Is there an MCP server for audio transcription with speaker diarization?
Yes. Gladia MCP supports speaker diarization as an option on its transcribe_pre_recorded tool, so “transcribe this call with speaker diarization” returns a transcript split by speaker. Our DER guide explains how diarization accuracy is measured.
Can Claude analyze diarized transcripts of sales calls?
Claude can analyze diarized sales call transcripts once Gladia MCP is connected. It sends the recording to Gladia with diarization turned on, reads the speaker-labeled transcript and can then summarize each speaker’s objections and agreed next steps across several calls in the same conversation.
Does Gladia MCP support real-time transcription?
Gladia MCP doesn’t start or stream live audio in its current release. It can check and delete live sessions created through the real-time API, which remains the right integration for voice agents.
Can Gladia MCP redact PII from transcripts?
Gladia MCP can turn on PII redaction for a pre-recorded job, replacing sensitive entities with markers or masks before the transcript reaches the assistant. As of October 2026, PII redaction costs 0.07€ per audio hour on top of the transcription rate. Our PII redaction explainer covers how it works.
