# Use your agents from Claude and other AI clients

> Call your org's Mindset agents from Claude, Claude Code, Cursor or Copilot in VS Code, without opening the Hub.

After this page, you can call your org's Mindset agents from your own AI client. You ask Claude to run the invoice exceptions agent on `INV-4471`, Claude hands it the work, and the answer comes back into the conversation you're already in.

## What this is

Mindset runs an MCP server for its users, named `mindset` in your client. **MCP** is a standard way for an AI client to see what another system offers and call it. Once you add Mindset's server to your client, your client can list the agents you can use and hand one of them a task.

You don't need to be an admin. If an agent is [available in the Hub](https://docs4.mindset.ai/docs/ams/where-your-agent-can-be-reached), you can reach it this way. There is no key or token to copy. You sign in with your usual Mindset account the first time your client connects.

Org admins also have a second server, `mindset-admin`, for building and governing Mindset itself. See [Build and govern Mindset from Claude](https://docs4.mindset.ai/docs/ams/build-and-govern-mindset-from-claude).

## Supported clients

The setup panel has instructions for Claude Code, Cursor, Copilot (VS Code), Claude Desktop and Claude.ai. ChatGPT isn't one of them. Each client's tab also has a manual fallback if you'd rather edit the configuration yourself.

## How to do it

1. **Open the Hub** and find **Use your agents from your own AI client**. If an admin sent you a setup message, its link takes you there.
2. **Pick the tool you work in.** The panel gives you something to copy.
3. **Paste it into your client.**
   - **Claude Code:** paste the prompt into a session. It runs `claude mcp add` with the server address, then asks you to run `/mcp`, pick `mindset` and sign in in the browser. It also adds a short "Using Mindset" note to the project's `CLAUDE.md`.
   - **Cursor** and **Copilot (VS Code):** paste the prompt into the client's agent. It writes the server into the project's MCP settings and a short note for the agent. In VS Code, reload, then run **MCP: List Servers**, pick `mindset` and choose **Start**.
   - **Claude.ai:** go to **Customize → Connectors → + → Add custom connector** and paste the address. On a Claude Team or Enterprise plan, an Owner adds the connector once in **Organization settings → Connectors**, and everyone else finds it under **Customize → Connectors** and clicks **Connect**.
   - **Claude Desktop:** add a custom connector in the app's connector settings and paste the address. On a Claude Team or Enterprise plan, an Owner adds the connector once and everyone else connects to it.
4. **Sign in** with your Mindset account when the client asks.
5. **Ask what you can call.** For example, "list the Mindset agents available to me". Your client calls `list_callable_agents` and shows the agents you can use, each with its description. An agent whose script needs start values lists them, and your client sends them on its first call.
6. **Call one.** "Use the invoice exceptions agent on INV-4471." Your client calls `call_agent`, the agent runs in Mindset, and the result comes back into your conversation.
7. **Keep talking to it, if you need to.** Each call returns a `contextId`. To carry on the same conversation, your client passes that `contextId` back and replays your earlier messages in `history`. A `contextId` on its own reaches the agent with no memory of the earlier turns.

![Claude's connector settings, with "Add custom connector" in the Add menu.](https://docs4.mindset.ai/images/docs/getting-started/claude-connectors.png)

## For admins: invite your people

In **Settings → Systems → Connect an AI client**, choose **For users**. The panel gives you a message to send by email or Slack, with **Copy the message**. It points people to their Hub page, where they pick their own client. Don't send them the administrators' address. Members can't reach it.

The panel's footer shows how many people have connected.

## What you should see

- **The agents you can reach.** A member sees the agents available to them in the Hub. An org admin sees every agent in the environment, including ones not on the Hub, because admins can reach any agent. Your own personal agents aren't listed.
- **An answer in your client's conversation.**

## Things to know

- **The agent uses its own permissions, not yours.** It reaches the [operations](https://docs4.mindset.ai/docs/ams/glossary) it was given, whichever route you came in by. Calling it from Claude doesn't make it able to do more or less than it can in the Hub.
- **A call can't be canceled once it starts.** If your client times out or you close the window, the agent carries on and finishes in Mindset.
- **Long calls come back as "working".** Mindset holds the call open for up to 30 seconds. If the agent is still going, the call returns the state `working` with no text. The agent hasn't stopped, but this server can't fetch the reply afterward. Don't call it again, because that starts a second run, and any writes in it happen twice. Check back later or ask an admin to look at the run.
- **A run has a time limit.** An agent called this way gets about four minutes per turn. Work that takes longer is better started by a [schedule](https://docs4.mindset.ai/docs/ams/triggers-and-schedules). See [Limits and run behavior](https://docs4.mindset.ai/docs/ams/limits-and-run-behavior).
- **Calls are rate limited.** Each caller has a rate cap, and the same delegation guards as every other route apply.
- **The address names one environment.** The server address includes the [environment](https://docs4.mindset.ai/docs/ams/environments), so you see only agents in that environment. If one you expect is missing, check which environment the address is for.
- **Writes that aren't approved yet are refused.** If one of the agent's write operations is still waiting for a person to approve it, the agent can't use it. Calling from Claude never approves anything. See [Approve what an agent can do](https://docs4.mindset.ai/docs/ams/approve-what-an-agent-can-do).

## When it does not work

- **The list is empty.** No agent in that environment is available to you in the Hub. Ask whoever owns the agent to tick **The Hub** for it.
- **The agent forgot what you said a turn ago.** Your client sent the `contextId` without `history`. Ask it to replay the earlier turns.
- **Your client reports a timeout or `working`.** The agent hasn't failed. Don't re-send. Wait, then check the result with whoever can read the run.
- **The agent says it can't do the step that changes something.** The write operation is waiting for approval. An admin approves it on the connection's **Operations** tab.

## You're done when

- You can list your agents from inside your client.
- You've called one and seen its answer.
- You know a `working` reply means wait, not retry.