MCP connectors

ReactLive has an MCP server: connect an AI assistant such as Claude to your workspace and run it from a chat. Create and configure events, approve and answer questions, run polls, manage the team and pull the report — by asking, in your own words. The assistant does exactly what you could do in the console, as you, and nothing more.

Before you start

• A ReactLive workspace you are a member of. The connection acts as you, in one workspace. What you can do through it is what your role allows in the console — an admin can do everything below; a moderator, the event side of it.
• An MCP-capable assistant. Claude.ai, Claude Desktop, Claude Code and Cursor are covered below. Any client that supports remote servers over streamable HTTP with OAuth sign-in will connect the same way.

The server

There is one server URL for everyone, whatever your workspace is called:

https://www.reactlive.com/api/v2/mcp

It is a remote server — nothing to download or run. Your workspace's name does not go into the address; you choose the workspace when you authorise the connection, and the connection then stays scoped to it.

Connect an assistant

Claude.ai and Claude Desktop

1. Open Connectors.

   In Claude, open Settings → Connectors and press Add custom connector.

2. Name it and paste the URL.

   Name: ReactLive. Remote MCP server URL: https://www.reactlive.com/api/v2/mcp. Leave the advanced fields empty — the server registers the client itself.

3. Connect and authorise.

   Press Connect. Claude opens ReactLive: sign in if you are not already, pick the workspace, press Authorize. You are returned to Claude with the connector on.

Claude Code

From a terminal, in the project you want to use it from (or with --scope user to have it everywhere):

claude mcp add --transport http reactlive https://www.reactlive.com/api/v2/mcp

Or add it by hand to .mcp.json in the project root:

{
  "mcpServers": {
    "reactlive": {
      "type": "http",
      "url": "https://www.reactlive.com/api/v2/mcp"
    }
  }
}

Then run /mcp inside Claude Code, choose reactlive and Authenticate — the same sign-in and workspace choice as above opens in your browser.

Cursor

Add the server to .cursor/mcp.json (project) or ~/.cursor/mcp.json (everywhere), then authorise it from Settings → MCP:

{
  "mcpServers": {
    "reactlive": {
      "url": "https://www.reactlive.com/api/v2/mcp"
    }
  }
}

Any other client

Give it the server URL as a remote (streamable HTTP) server. ReactLive supports the standard discovery chain — OAuth 2.0 with PKCE, dynamic client registration and refresh tokens, advertised at https://www.reactlive.com/.well-known/oauth-authorization-server — so a client that follows the MCP specification needs no client ID or secret from you.

Authorising, and choosing the workspace

The first time a client connects, ReactLive opens an authorisation page in your browser. If you are not signed in it asks you to sign in first. The page lists every workspace you belong to; choose the one this assistant should act on and press Authorize. What the client receives is a token for that workspace only, with read and write access within it.

Switching workspace

Disconnect the server in your assistant and connect it again — the authorisation page comes back and you can choose a different workspace. One connector is one workspace; add the server twice under two names if you work across two.

Taking access away

Disconnect it in the assistant. The token stops working; nothing else about your account changes.

Who it acts as

You. Every event it creates, every approval and every answer is recorded against your account, the same as if you had done it in the console.

What you can do

You ask in plain language and the assistant picks the tool. Behind the scenes the server exposes about thirty tools; this is what they cover, in console terms.

The assistant always starts by listing your events so that it is working on the right one; it will ask when a request is ambiguous ("which town hall?").

Things to ask

Some prompts that work as written. Replace the names with your own; the assistant looks up the IDs.

Set up an event

"Create a town hall called 'Q3 All-Hands' on 24 September, 2pm to 3:30pm Dublin time. Hold questions for approval and turn on upvotes."

Bring in the team

"Add maria@example.com and tom@example.com as moderators on Q3 All-Hands." If someone is not yet a member of the workspace the assistant asks before creating them.

Feed the AI

Knowledge sources are added under Settings → AI knowledge in the console — the connector does not upload files. It can tell you what the event's settings say, so "Is the Answer agent on for Q3 All-Hands?" works.

Moderate during the event

"Show me the questions waiting for approval on Q3 All-Hands" — then "approve all of them except the one about parking". Or "Pin the question about the new office to the top."

Run a poll

"Add a poll to Q3 All-Hands: 'Which should we prioritise next quarter?' with options Hiring, Tooling, Customer visits, Training."

After the event

"Give me the numbers and a summary for last week's Q3 All-Hands" — the counts, the sentiment and the takeaways the Report agent already wrote, so it answers instantly. "Write a fresh summary of the last thirty minutes" runs the AI again over that window. "How much did we spend on AI in August?" gives the per-agent breakdown.

Permissions and safety

• Same rules as the console. The server checks every call against your role in the workspace you authorised. It cannot reach other workspaces, even ones you belong to, without a separate authorisation.
• Everything is on the record. Each tool call is written to the workspace's audit log, attributed to your account and naming the tool — so a colleague can always see that you approved those questions, via an assistant.
• Destructive actions ask twice. Deleting an event, a post or a category needs an explicit confirmation flag the assistant can only set after you have confirmed; there is no wording that skips it. Prefer un-approving over deleting — it is reversible.
• The assistant can be wrong. It chooses which tool to call from what you typed. Read back what it says it did — especially before it publishes — and keep the console open beside it during a live event.

Rate limits

Each connection may make 60 requests a minute. Beyond that the server answers 429 Too Many Requests with an X-Rate-Limit-Retry-After-Seconds header, and a well-behaved assistant waits and retries on its own. If you keep hitting the limit — bulk approvals over a large backlog, say — ask for the work in smaller batches.

Tool catalogue

A public, unauthenticated catalogue of every tool — names, descriptions, input schemas and annotations — is at either of these addresses:

https://www.reactlive.com/api/v2/mcp/tools.json
https://www.reactlive.com/.well-known/mcp.json

Both return the same document, combining the initialize and tools/list shapes from the MCP specification. It is useful for seeing what the server can do before you connect anything, and for MCP directories that index servers by their tool surface. Reading the catalogue needs no sign-in; calling a tool always does. The address is also advertised as catalog_endpoint in the OAuth server metadata, so clients that follow the discovery chain find it on their own.

MCP and the AI agents

These are two different things that happen to both involve AI. The agents — Engage, Answer, Protect, Report — are ReactLive working inside the event, on the room's behalf, at the autonomy you set. An MCP connection is you, working on ReactLive from your own assistant, with your own permissions. Nothing you do over MCP changes an agent's level; and an agent never uses your connection.

Troubleshooting

Still stuck? Tell us which assistant you are using and what you asked it.