MeetStream

MeetStream MCP Server

Give Claude - or any MCP-compatible client - direct access to the MeetStream meeting-bot API. Add one URL, sign in, and your assistant can run the full bot lifecycle (create, record, transcribe, summarize, interact live, manage calendar auto-join) through 19 tools.

https://mcp.meetstream.ai/mcp

MeetStream hosts the server. There is nothing to install or run on your machine, and it is always the current version.

Quick start

1. Add https://mcp.meetstream.ai/mcp to your AI tool. Do not add a key.
2. Click Connect (in Claude Code: run /mcp, then choose Authenticate). Your browser opens app.meetstream.ai.
3. Sign in to MeetStream, then click Allow. MeetStream creates an API key for this connection. You do not see or copy the key.
4. Ask your AI tool: *"list my meetstream bots"*. A list, or an empty list, means that you are connected.

The exact steps for each tool are in Connect.

Which URL to use

Your AI tool ──HTTPS + OAuth sign-in──► mcp.meetstream.ai ──► api.meetstream.ai

The old npm package @meetstream/mcp (npx @meetstream/mcp) is frozen at 0.3.0 and no longer maintained. It predates the current create_bot defaults and fixes. Use the hosted URL instead.

Connect

Add the URL with no key. Your client finds MeetStream's sign-in, opens app.meetstream.ai in your browser, and you click Allow. The dashboard creates an API key for that connection, named MCP · · · . To disconnect, delete that key on the API keys page.

This is the page that opens in your browser. If you are not signed in to MeetStream, you sign in first, then you see this page:

The page shows the name of the app that asked to connect. If you do not recognize the app, click Cancel. The sign-in link is valid for 15 minutes.

Claude Code

claude mcp add --transport http meetstream https://mcp.meetstream.ai/mcp

then run /mcp in Claude Code and choose Authenticate

Claude (web and desktop)

Customize → Connectors → + → Add custom connector. Name it meetstream, set the URL to https://mcp.meetstream.ai/mcp, leave the OAuth fields blank, and click Connect.

Cursor

~/.cursor/mcp.json (or install the MeetStream Cursor plugin), then click Connect next to meetstream in Cursor Settings → MCP:

{

"mcpServers": { "meetstream": { "url": "https://mcp.meetstream.ai/mcp" } } }

VS Code

.vscode/mcp.json, then start the server and sign in when VS Code asks:

{

"servers": { "meetstream": { "type": "http", "url": "https://mcp.meetstream.ai/mcp" } } }

Windsurf

~/.codeium/windsurf/mcp_config.json. Cascade asks you to sign in on first use. If your Windsurf version cannot finish the sign-in, use an API key header (below):

{

"mcpServers": { "meetstream": { "serverUrl": "https://mcp.meetstream.ai/mcp" } } }

Codex

~/.codex/config.toml, then run codex mcp login meetstream:

[mcp_servers.meetstream]

url = "https://mcp.meetstream.ai/mcp"

Any other MCP client

Point it at https://mcp.meetstream.ai/mcp as a remote (Streamable HTTP) server. Clients that support MCP authorization sign in by themselves: protected resource metadata (RFC 9728), authorization server metadata (RFC 8414), dynamic client registration (RFC 7591) and the authorization code flow with PKCE. Access tokens last an hour and refresh automatically.

Or use an API key

For clients without OAuth, or for scripts and CI, create a key at [app.meetstream.ai/api-keys](https://app.meetstream.ai/api-keys) and send it as a header:

{

"mcpServers": { "meetstream": { "url": "https://mcp.meetstream.ai/mcp", "headers": { "Authorization": "Bearer ms_YOUR_API_KEY" } } } }

• Claude Code: claude mcp add --transport http meetstream https://mcp.meetstream.ai/mcp --header "Authorization: Bearer $MEETSTREAM_API_KEY"
• Codex: add bearer_token_env_var = "MEETSTREAM_API_KEY" to the [mcp_servers.meetstream] block.
• X-MeetStream-Api-Key: works too. For connector UIs that cannot set headers, https://mcp.meetstream.ai/mcp?key= . The server never logs query strings, but a key in a URL can end up in client-side history, so prefer OAuth or a header.

Older clients (HTTP+SSE)

/mcp is the main endpoint, and its replies already stream as server-sent events (text/event-stream). Clients that only speak the older 2024-11-05 HTTP+SSE transport connect to https://mcp.meetstream.ai/sse with the same sign-in or key options. The stream's first event names the per-session message endpoint; the key never appears in it.

Check it is connected

Ask your client something like *"list my meetstream bots"* - if it returns data (or a clean empty list) instead of an error, you are connected.

Then just ask

"Join my standup at https://meet.google.com/abc-defg-hij, record it with Deepgram transcription, and give me the transcript when it's done."

"List my bots from today and give me the AI summary of the 3pm customer call."

"Send 'we'll follow up by email' into the meeting the bot is in, then show them our logo."

"Schedule a bot for tomorrow's board meeting on my calendar."

Full capability list - 19 tools

Bot lifecycle

Transcription

Media & meeting data

Live meeting interaction

Calendar

Reference

What makes this different from just reading the docs

Every tool description and the webhook_events_guide bake in live-verified ground truth, confirmed against real production bot runs, that the public API docs currently get wrong:

• Every delivery carries `event`, and most also carry bot_event with the specific name. Read event to route, bot_event for detail.
• Terminals are two-layer: every ending arrives once with event: "bot.stopped", and bot_event says why (bot.stopped, bot.kicked, bot.notallowed, bot.denied, bot.failed). Lobby timeouts, denials and failures carry status_code: 500. Branch on bot_event, not bot_status - a kick and a clean exit both report Stopped.
• Streaming-only transcription providers (deepgram_streaming, assemblyai_streaming, jigsawstack_streaming, meetstream_streaming, meeting_captions) produce no post-call transcript, so transcription.processed never fires. bot.done still does. A post-call transcript fetch for them returns 202 indefinitely, so cap your polling.
• `transcript_id` is never in a webhook payload - get_transcript resolves it for you automatically instead of making the model guess.
• Safe defaults everywhere: automatic_leave timeouts on every create_bot call, and recording_permission_denied_timeout floored at 60 (the API rejects lower values with a 400).

How it works

The server holds no API key of its own: each request (or SSE session) carries the caller's key or an OAuth token that wraps it. A request with no credentials gets a 401 with setup instructions and a WWW-Authenticate header pointing OAuth clients at the sign-in flow. Streamable HTTP is stateless, so the service scales horizontally with