Speak AI MCP server, connect Claude, ChatGPT, and more

The Speak AI MCP server connects Claude, ChatGPT, Cursor, VS Code, Muse, and other Model Context Protocol clients to a Speak AI workspace. It runs at https://api.speakai.co/v1/mcp and exposes 185 tools, 5 resources, and 3 prompts covering media, transcripts, AI insights, folders, recorders, automations, dashboards, webhooks, exports, and voice agents.

Those counts are for server version 1.36.0, last verified 2026-10-15. Eleven of the tools are registered only when the server runs locally over stdio: upload_local_file, which reads a file from the machine the server runs on, and the ten voice agent test-run tools, whose execution engine is not live yet. The remote endpoint offers the other 174.

The same package that implements the server, @speakai/mcp-server, is also a Node library. If you want to call Speak AI from your own code instead of through a connected AI client, see the Node SDK.

What can an agent do once it’s connected?

Once connected, an agent can search recordings, read transcripts and AI insights, create folders and clips, run automations, and schedule a meeting assistant, all without you leaving the chat. You describe what you want, and the agent picks the right tools.

Examples of what people ask once the server is connected:

• “Find the last 10 customer interviews that mention pricing, group the feedback by theme, and cite the source recordings.”
• “Summarize this week’s team meetings into decisions, action items, owners, and unresolved risks.”
• “Join my 2pm Zoom call, then send me a summary with action items.”
• “Find a strong 30-second highlight from the latest webinar, create a clip, and export captions.”

Your recordings stay in your Speak AI workspace. Claude, ChatGPT and Muse only see the specific data a request asks for, and none of them copies or stores your media.

Which client are you setting up?

Connecting takes one URL paste and one permission click for most AI tools. Paste the address above into your client’s connector settings, then approve the popup that opens.

Pick yours for the steps.

How does authentication work?

The MCP server accepts two authentication methods: OAuth 2.1 with Dynamic Client Registration for the one-click connect flow, or a Speak AI API key passed as a Bearer token for manual setups, stdio mode, and the CLI.

See Authentication for how to get a key, what the REST token flow looks like, rate limits, and the tool error format.

Where’s the full tool reference?

Every tool, resource, and prompt the server exposes is documented in the tool reference, grouped by area: media, AI chat, folders and views, recorders, automations, clips, custom fields, webhooks, users and teams, dashboards, the meeting assistant, media embeds, text notes, exports, search, and voice agents.

You don’t need to memorize tool names. Claude, ChatGPT and Muse pick the right ones based on what you ask.

Is there a Node SDK too?

The same package that runs the MCP server, @speakai/mcp-server, is also a library you can import directly into Node code. Use it when you’re embedding Speak AI’s tools inside your own service instead of connecting through an AI client.

See Node SDK for the exported functions and when to reach for them.

What does connecting actually grant access to?

Both methods, the permission popup and a pasted API key, grant the AI assistant read and write access to your whole Speak AI workspace. That includes viewing, modifying, and deleting recordings, folders, and clips.

Access is not continuous. Claude, ChatGPT or Muse reads specific data only when it calls a tool to answer something you asked, and your recordings stay in your Speak AI workspace.

Which workspace does the connector act on?

A Speak AI account belongs to exactly one workspace, and every credential is bound to that same workspace, so the connector always acts on that one workspace and there is nothing to disambiguate. If you work across two workspaces, connect the account that belongs to each one separately.

What can the connector do in each area?

Read means it can list or fetch. Create means it can add something new. Modify means it can change something that already exists. Delete means it can remove it.

Exports read your media and hand back a file to download, so they create nothing in your workspace. Users and teams covers user groups only: the connector can list workspace members but it cannot create, change or remove a user. Voice agents covers agents, the questions they ask callers and their test suites. Over the connector, testing stops at the suite: the ten tools that queue, pause, cancel and read a test run are stdio-only, so a connected client can build or generate a suite but not run it.

Destructive actions

Thirteen tools delete data. MCP marks all of them as destructive and most clients prompt before running one, but ask your assistant to confirm with you before it deletes anything.

Deleting is not the only thing marked that way. A tool that overwrites a record, spends credits, fires your workspace’s webhooks, runs a connected app’s action or creates a public link carries the same marker, so your client may ask you to confirm those too.

Five of them are worth knowing by name, because what they remove is wider than it sounds:

The other eight are delete_chat_conversation, delete_automation, delete_webhook, delete_user_group, delete_scheduled_assistant, delete_dashboard, delete_voice_question, and delete_voice_agent_resource. Deleting a dashboard is a soft delete, and it also deactivates that dashboard’s public share link. Deleting a voice question keeps its template, and deleting a voice agent resource is a soft delete.

To disconnect, remove the connector inside Claude, ChatGPT or Muse, or revoke a one-click connection at api.speakai.co/v1/oauth/connections.

What do other people see when the meeting assistant joins a call?

The meeting assistant joins Zoom, Google Meet, Microsoft Teams and Webex calls, and it never joins one silently. It appears in the participant list like any other attendee, with its own name and avatar tile that everyone on the call can see, and it starts recording the moment it joins. It is there because you sent it, either by scheduling it for a specific call or by setting an auto-join rule for a calendar you connected, and anyone on the call can remove it the way they would remove any participant, which stops the recording immediately. You can also pause the recording over a confidential stretch and resume it afterwards, and nothing said while it is paused reaches the transcript. What it captures lands in your Speak AI workspace as a recording with a speaker-labeled transcript and a summary, routed to the folder you chose, and it stays there under your workspace’s normal sharing rules until you share it. Recording laws differ by region, so tell attendees they are being recorded and get their consent where it is required.

How is this different from AI chat inside Speak AI?

Speak AI’s built-in AI Chat answers questions inside Speak. The MCP server lets you ask the same questions from Claude, ChatGPT or Muse, where you may already be drafting an email, planning content, or writing code. Use whichever fits the task.

What if the connection keeps failing?

A single failure mid-conversation is usually transient, so try the same prompt again. If it persists, reinstall the connector: revoke the one-click connection at api.speakai.co/v1/oauth/connections, then connect again.

If your company blocks custom connectors, that is an admin policy rather than a fault. Some Claude Team and Enterprise workspaces restrict them, and your workspace admin needs to allow https://api.speakai.co/v1/mcp.

Related guides

• Which client are you setting up?
• Authentication
• Node SDK
• Tool reference