Sonix MCP Server

What is an MCP server?

An MCP server is the service side of the Model Context Protocol: it exposes a defined set of tools an AI assistant can call, with your authorization. The Model Context Protocol (MCP) is an open standard that connects AI assistants to outside tools and data. Anthropic introduced it in late 2024, and it has since been adopted across the AI ecosystem, including by OpenAI and GitHub. It lets an assistant such as Claude or ChatGPT discover what a service can do, then call those capabilities on your behalf.

Before MCP, connecting an AI assistant to an app meant a custom integration for every pairing: one for Claude, another for ChatGPT, another for your editor. MCP replaces that with one protocol. A service runs one MCP server, and every compatible client can connect to it.

That matters more for transcripts than for most data. A transcript library is exactly the context an assistant needs and cannot reach on its own: hours of meetings, interviews and calls, kept in another app. Pasting a transcript into a chat works once. It does not work across a 40-file interview project.

The Sonix MCP server closes that gap. It is a remote MCP server, hosted by Sonix, so there is nothing to install and nothing to run on your machine. Point your client at one URL, approve the connection in your browser, and your transcript library becomes something your AI can work with directly.

What your AI can do with Sonix

Sonix runs a hosted MCP server at https://api.sonix.ai/mcp that lets any compatible client read your media library, transcripts and account details on your behalf.

Today's surface is deliberately read-only: browsing media, fetching transcripts and exporting files. Write tools for creating transcriptions, translations and edits are on the roadmap.

Browse your media in plain English

No more copying transcript links or downloading files just to ask a question. Tell your assistant what you want, whether that is finding a recording, pulling a quote or summarizing an interview, and it uses Sonix on your behalf.

Transcripts straight into the conversation

Bring a speaker-labeled transcript into the chat for questions, summaries, sentiment, entity extraction or any other analysis a language model can do. No copy and paste, no manual download step.

Exports without leaving your tool

Ask for a clean text transcript, SRT or VTT subtitles, or structured JSON. Sonix creates a short-lived download link your AI can save or pass along, so large files stay out of the conversation and do not use up tokens.

A media library inside your chat

In clients that support MCP Apps, including Claude and ChatGPT, the media list arrives as an interactive Sonix library rather than plain text. Browse your recordings, open a transcript and export it from inside the chat.

Who can connect

The MCP server is included with every paid Sonix plan at no extra cost; trials and free accounts cannot connect MCP clients. Your role on the account must be owner or producer to authorize a connection, and members and guests cannot. That is deliberate: a member-level user can never hand account-wide transcript access to an outside AI tool by accident.

Set up the Sonix MCP server in three steps

Setup takes about a minute: a paid plan, one server URL in your client, and one approval in your browser.

1. Have a paid Sonix account

The MCP server ships with every paid plan, with no separate add-on and no extra price. New to Sonix? Start a free trial, then move to any paid plan to turn on MCP.

2. Point your client at the server URL

Every client connects to the same address:

https://api.sonix.ai/mcp

For Claude Code, one command does it:

claude mcp add --transport http sonix https://api.sonix.ai/mcp

For Claude Desktop, edit claude_desktop_config.json (Settings > Developer > Edit Config) and bridge through mcp-remote like this:

{
  "mcpServers": {
    "sonix": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://api.sonix.ai/mcp"]
    }
  }
}

ChatGPT, Cursor, Codex, Windsurf and VS Code each keep their configuration somewhere different, and the setup snippets in our docs cover every one of them, ready to copy. Any other client that speaks the current MCP specification over Streamable HTTP with OAuth 2.1 can connect to the URL directly.

3. Approve the connection, then ask

The first time your client connects, Sonix opens in your browser. Sign in, check which tool is asking for access, and click Allow. Then talk to your assistant in plain English: "Pull up my interview with Dr. Patel and summarize the takeaways." "Generate SRT captions for my latest podcast." Your AI does the rest.

Five MCP server examples from real transcript work

An MCP connection is only as useful as the questions you ask through it, and these five workflows run today on the server's four read-only tools.

Mine a quarter of calls for one topic

Ask Claude: "Go through last quarter's customer calls and find every mention of pricing objections." Your assistant lists your recent media, reads each transcript and comes back with the quotes, who said them and which calls they came from. An afternoon of searching through downloads becomes one prompt.

Turn a week of meetings into a digest

"Summarize every meeting from last week and flag action items with owners." The assistant pulls each meeting transcript into context and writes one digest. Because transcripts are speaker-labeled, action items come back attached to names, not just topics.

Pull exact, attributed quotes for an article

For journalists and researchers: "Find where the CFO discussed the acquisition in my March interviews and give me her exact words." The assistant quotes from the source transcript instead of paraphrasing from memory, which is the difference between checkable and not. Need timing for a fact-check? Ask for a text export that includes timestamps.

Audit interviews against a theme

For UX and qualitative researchers: "Read all twelve usability interviews and tell me which participants ran into the onboarding confusion we tagged in the pilot." The assistant works through the set methodically, with evidence attached. When a transcript is larger than the 50 KB inline limit, the server points the assistant to an export instead, so long sessions keep working.

Feed an agent that drafts the follow-ups

Connect Sonix to an agent in Claude Code or Cursor and go beyond questions and answers. After each sales call, the agent reads the transcript, drafts a recap email with commitments and next steps, and files it for review. The MCP server is the retrieval layer; your agent framework decides what happens next.

Read-only by design

The Sonix MCP server was scoped around the first question every security reviewer asks: what can a connected AI tool do if it misbehaves?

Read-only access

Every current tool reads and none writes. A connected client cannot edit transcripts, delete media or change account settings. When write tools ship, they will be documented first and arrive as tools you can see, not silent changes to what an existing connection can do.

OAuth 2.1, not shared keys

Connections use OAuth 2.1, the current version of the open standard for delegated access. You approve each client by name in your browser, access tokens expire after an hour and refresh automatically, and you can revoke any connection from your Sonix profile at any time. Your REST API key is never involved: the two surfaces have independent authentication.

Role-gated authorization

Only account owners and producers can authorize an MCP connection. Members and guests cannot, so account-wide transcript access stays in the hands of your admins.

The same controls as the rest of Sonix

MCP access inherits the rest of the Sonix security posture: SOC 2 Type II certification, GDPR compliance, HIPAA-compliant plans, AES-256 encryption and a strict no-AI-training policy. Sonix never sells, shares or trains on your data, and connecting an MCP client does not change that.

MCP server, REST API or CLI: which one you need

Sonix gives developers three ways into the same platform, and they complement each other.

The MCP server, when a person is driving an AI assistant

It is conversational, needs no code and shines at retrieval and analysis: find, read, summarize, export. If the request starts with "ask Claude to", this is the surface to use.

The REST API, when you build software

Uploading media programmatically, starting transcription from your pipeline, receiving webhooks when jobs finish and shipping transcription inside your own product all belong to the Sonix REST API. It has full write access and predictable behavior, which is exactly what production systems need and what a language-model loop is not.

The Sonix CLI, when you want the API from a terminal or a script

sonix is the official command-line client: it wraps the REST endpoints, installs with npm install -g @sonix-ai/cli and authenticates with your existing API key. It serves shell scripts and scheduled jobs, and it also serves AI agents that would rather run a command than spend tokens on tool calls. The CLI announcement has the details.

A common production setup uses all three: the REST API ingests recordings automatically, the CLI handles batch exports in CI, and the MCP server lets the team question the resulting library from Claude.