Connect Raconte to Claude Tag

Claude Tag puts an AI teammate in Slack: you tag @Claude in a channel and it runs a task with the tools an admin connected for it. Raconte is one of those tools. Once connected, @Claude can launch voice interviews with real people, then read back the transcripts, summaries and sentiment without leaving the thread.

This guide shows how to make that connection. For why you would want it, start with Putting real human answers into Claude Tag.

Before you start

• You are a Claude Tag admin (Primary Owner or Owner) of the Slack workspace, and Claude Tag is already paired. Only an admin can edit access bundles.
• You have a Raconte API key. Create it in the app under Settings → API. It is scoped to one organization: every call only touches that organization.

1. Get the MCP endpoint

Raconte exposes an MCP server over Streamable HTTP. The endpoint carries the key in the URL:

https://api.raconte.ai/api/mcp?api_key=YOUR_API_KEY

A ready-to-paste configuration is also available in the app under Settings → MCP. The MCP page lists the tools @Claude gets once connected (on interviews and invitations) with their parameters, and the Using the MCP server guide shows examples phrased in plain language.

2. Add Raconte to the channel’s access bundle

In Claude Tag, each Slack channel has its own access bundle: a named set of credentials and tools that @Claude uses in that channel, and nowhere else.

1. Open the Claude Tag administration and pick the channel that needs it most (say #content).
2. In its access bundle, add a Streamable HTTP MCP server.
3. Paste the Raconte endpoint with your key.

Channel scoping handles the rest: the Raconte key only lives in that channel. The #sales channel has no access until you add it there explicitly.

3. Set scope and spending

First, keep the access bundle on the channels that actually have a reason to interview people. On the budget side, hold two distinct things in mind:

• The channel’s Claude Tag spending limit (under an organization-wide cap) bounds what the agent itself spends: its reasoning, its drafting, its tool calls. Set it by how you use @Claude in that channel, not by your interviews.
• The cost of the interviews is billed separately by Raconte, per minute of active speech, with free monthly minutes per organization. It does not count against the Claude Tag limit and is tracked from your Raconte account.

4. Tag @Claude

Once Raconte is in the bundle, delegate a task in plain language:

@Claude, with Raconte, interview our data expert and three customers
who use the reports daily, then propose an article angle and the most
telling quotes.

Saying “with Raconte” helps @Claude pick the right tool when the channel has several. It creates the interviews, sends the invitations, then comes back with the material (an angle, the best verbatims) once the interviews are done. The next section explains how it knows they are.

How @Claude knows an interview is done

The MCP server is pull-based: Raconte does not push anything into Slack on its own. So @Claude has to read the state with the interview-read tool. Two ways, depending on whether you want polling or events:

• @Claude reschedules a check for itself. Claude Tag can schedule follow-on work for itself (“check back in two hours”, “follow up tomorrow morning”). It then re-reads the interview until the invitations turn completed, and posts the synthesis. That is what “later” means in the thread below: @Claude scheduled a re-check, it is not waiting passively.
• A webhook pushes the information into the channel. To avoid any repeated checking, wire Raconte’s interview-completed webhook to post a message into the channel (directly, or through an automation like n8n or Make). In ambient mode, @Claude sees that message arrive and carries on. There it is the event that triggers the next step, not polling.

Ambient mode also enables scheduled routines (re-interview every quarter, chase unfinished interviews), but that is not required for the basic case.

Example Slack thread

Ambient mode off, so triggered by a single tag, with @Claude handling its own return:

@you     @Claude, with Raconte, interview our data expert and three
         customers about how they use the reports, then propose an
         article angle and the best verbatims.

@Claude  On it. I am creating the Raconte interviews and sending the
         invitations. I am scheduling a check in two hours and will
         come back as soon as answers are in.

         … (two hours later, @Claude re-checks the state) …

@Claude  4 interviews completed. The angle that emerges: teams do not
         use the reports to track, but to decide in meetings. Three
         strong customer verbatims on that, including a concrete
         anecdote. Angle proposal and a first draft in the thread.

Guardrails

• Sample size. Ask for a volume of interviews and do not conclude on three answers.
• Interview prompt. An open question (“walk me through how you use the reports”) collects richer material than a leading one (“are the reports useful to you?”).
• A human in the loop. @Claude’s synthesis is a starting point. The full transcripts are there so a human can check before turning it into a decision.

See the MCP server reference