---
title: "Agent lab (MCP): train your agent from your AI tool"
description: "Connect Claude Code or any MCP client to gptagent: create, edit, rehearse, check and publish your own AI agents. Publishing waits for a passed quality check."
url: https://gptagent.ai/mcp/
lang: en
---

Product / Agent lab (MCP)

# Train your agent from your own AI tool.

The agent lab is a public MCP server. Connect Claude Code or any other MCP client with your access key, and work on your gptagent agents in plain conversation: create one from a template, rewrite its instructions, rehearse a dialog, run the quality check and publish. No developer needed.

[Start free](https://app.gptagent.ai/signup?lang=en&utm_source=site&utm_medium=cta&utm_content=mcp) [How to connect](#ml-connect)

How it works

## One loop: create, edit, rehearse, check, publish.

The same steps every time. Your AI tool calls the lab's tools for you, and you see every result in the conversation.

1.  ### Create
    
    Start a new agent from a ready template, or pick one you already have.
    
2.  ### Edit
    
    Open a training session and rewrite the agent's instruction and knowledge files. Add the rules it must follow and must never break.
    
3.  ### Rehearse
    
    Play a practice dialog between a simulated customer and your agent, in chat or voice, and read the full transcript.
    
4.  ### Check
    
    Save your changes as a new version and run the quality check against your test set. It takes a few minutes and ends with passed or failed.
    
5.  ### Publish
    
    Publish the version that passed. Or show it to part of your traffic first, as a Challenger next to your Primary version.
    

Safe by design

## Your customers only meet versions that passed.

### Publishing waits for the quality check

A version goes live only when a finished quality check of that exact version passed. A missing, unfinished, failed or other-version check means no publish. No flag skips this.

### Only your company

Your company comes from your access key. No tool takes a company as a parameter, so you see and change only your own agents.

### The key carries your rights, not more

When you create a key, you choose its roles. A key never gets more rights than you have. Publishing, rollback and accepting a version need the right to publish.

Connect

## Connect your MCP client in three steps.

1.  In the gptagent console, open Channels → Access keys and create an access key. Copy it right away: it is shown only once.
2.  In your MCP client, add a new server with the address below.
3.  Add the authorization header with your key. The tools appear after the client connects.

<table data-astro-cid-7y3phznp=""><tbody data-astro-cid-7y3phznp=""><tr data-astro-cid-7y3phznp=""><th scope="row" data-astro-cid-7y3phznp="">Server address</th><td data-astro-cid-7y3phznp="">https://app.gptagent.ai/mcp/ (keep the trailing slash)</td></tr><tr data-astro-cid-7y3phznp=""><th scope="row" data-astro-cid-7y3phznp="">Transport</th><td data-astro-cid-7y3phznp="">Streamable HTTP</td></tr><tr data-astro-cid-7y3phznp=""><th scope="row" data-astro-cid-7y3phznp="">Authorization</th><td data-astro-cid-7y3phznp="">Header Authorization: Bearer &lt;your access key&gt;</td></tr><tr data-astro-cid-7y3phznp=""><th scope="row" data-astro-cid-7y3phznp="">Company</th><td data-astro-cid-7y3phznp="">Taken from the key automatically, you do not enter it</td></tr></tbody></table>

Example: Claude Code

```
claude mcp add --transport http gptagent \
  https://app.gptagent.ai/mcp/ \
  --header "Authorization: Bearer your-access-key"
```

Example: MCP client config file

```
{
  "mcpServers": {
    "gptagent": {
      "type": "http",
      "url": "https://app.gptagent.ai/mcp/",
      "headers": {
        "Authorization": "Bearer your-access-key"
      }
    }
  }
}
```

The access key is a password. Do not paste it into chats or send it to anyone. Replace your-access-key in the examples with your own key. You can revoke a key in the console at any time; connections that use it stop working right away.

What your AI tool can do

## The tools, in plain words.

### Agents

List your agents, open one, create a new one from a template, rename or delete it. Your company's main agent (default) cannot be deleted.

-   `list_agents`
-   `get_agent`
-   `create_agent`
-   `rename_agent`
-   `delete_agent`

### Training and files

Open a training session, read and rewrite the instruction and knowledge files, set the rules of the test set, choose how the agent picks its voice.

-   `start_session`
-   `read_instruction`
-   `write_instruction`
-   `set_gate`
-   `set_voice_policy`

### Memory

Short facts the agent remembers between conversations, with an optional expiry such as "until the end of the week".

-   `list_memory`
-   `add_memory`
-   `update_memory`
-   `delete_memory`

### Versions

Every change is saved as a version. Look at older versions and what changed, undo the last one, keep several draft lines next to the live agent.

-   `commit_candidate`
-   `list_commits`
-   `list_branches`
-   `fork_branch`
-   `training_log`

### Rehearsal and quality check

Play a practice dialog in chat or voice, score a transcript, run the full quality check and get passed or failed.

-   `run_dialog`
-   `analyze_dialog`
-   `run_qa`
-   `get_run_qa`

### A/B: Primary and Challenger

See your results by version. Send part of the traffic to a Challenger while the Primary version keeps the rest, then make the Challenger the new Primary or stop the test.

-   `get_kpi`
-   `get_experiment`
-   `promote_challenger`
-   `promote_champion`
-   `retire_challenger`

### Publishing

Publish a version that passed the check, or roll back to the previously published one.

-   `promote`
-   `rollback`
-   `adopt_candidate`

Good to know

## What the lab does and does not do.

-   Phone, chat and messenger channels are set up in the console, not here. The lab prepares and checks the agent; a rehearsal only simulates a channel.
-   Rehearsals and quality checks never touch the live agent: customers keep talking to the published version.
-   When a step is refused, the tool answers with a clear error, for example no right to publish or the check did not pass. Your key never appears in an answer.
-   Requests are rate-limited per key. If you send too many in a short time, wait a moment and retry.
-   Access keys can be created by company admins and developers.

## Create an account and connect your AI tool.

Sign up, create an access key in the console and train your first agent from Claude Code or any MCP client.

[Start free](https://app.gptagent.ai/signup?lang=en&utm_source=site&utm_medium=cta&utm_content=final)
