# Vatio in Claude and ChatGPT

**Build and run your agent from a chat.** Vatio is a remote MCP server at
`https://vatio.ai/mcp`. Add it to claude.ai, Claude Code, ChatGPT or any other
MCP client, sign in with your Vatio account, and ask for what you want: "make
an agent for my store's customers", "why did it answer that yesterday?",
"publish it". There is nothing to install, and it works from the browser, the
desktop apps and your phone.

It does what the CLI and the inbox do, from wherever you already talk to
Claude or ChatGPT. When your agent's files also live in your app's repository,
see [Your files and the MCP](#your-files-and-the-mcp).

> Looking for the agent inside your app? The MCP makes and runs the agent; to
> put it in your app (web widget or iOS, Android and Flutter SDK), run
> `npx @vatio-ai/skills` in the project and `/vatio-init`, which adds the chat
> screen in your app's style.

## Connect

The address is the same everywhere:

```text
https://vatio.ai/mcp
```

The first time, your client opens vatio.ai to sign in (email, Google or
GitHub, plus your second step if you set one up). Vatio shows which app is
asking and where it will send you back; approve it, and you are connected.

- **claude.ai** and the Claude apps: add a custom connector with that URL
  (Settings → Connectors).
- **Claude Code**: `claude mcp add --transport http vatio https://vatio.ai/mcp`,
  then run `/mcp` to sign in.
- **ChatGPT**: custom connectors need developer mode on your plan (Settings →
  Apps & Connectors → Advanced). Then create a connector with that URL.
- **Any other client** that speaks MCP over Streamable HTTP with OAuth, such
  as Cursor: point it at the URL.

Each connection is a session of your account. You see it, named after the
app, under **Account → Sessions** in the [inbox](https://vatio.ai/docs/channels/inbox), and you can
disconnect it there at any time. Signing the CLI in again does not disconnect
it.

## What you can ask for

The tools act as you, on the workspaces you can reach:

| Area | Tools | What they do |
| --- | --- | --- |
| Account | `list_workspaces`, `create_workspace` | The workspaces you reach and your role in each; a new one for a new agent |
| The agent | `get_agent`, `update_agent`, `set_tool`, `remove_tool` | Read the agent as the files a repository would hold; change its name, instructions, personality, tools, knowledge or [links](https://vatio.ai/docs/links); add or replace a [tool](https://vatio.ai/docs/tools) as its whole `tools/NAME.yml` |
| Knowledge | `list_knowledge`, `read_knowledge_entry`, `write_knowledge_entry`, `delete_knowledge_entry`, `create_knowledge_base` | The [knowledge bases](https://vatio.ai/docs/knowledge) the agent answers from |
| Trying it | `chat_with_agent`, `run_tests`, `get_test_results` | Talk to the agent as a tester and see the tools it called; replay every reported reply against it |
| Shipping | `get_status`, `publish`, `rollback` | What preview and live run; take preview live, through the same [publish gate](https://vatio.ai/docs/cli/deploy#the-publish-gate); put live back |
| Conversations | `list_conversations`, `get_conversation`, `list_reports`, `report_reply` | Real conversations (personal data redacted), one in full with its tool calls, the replies your team reported, and reporting one yourself |
| Privacy requests | `export_contact`, `erase_contact` | Everything kept about one person as JSON, or delete it all for good (`confirm_contact_id` must repeat the id). Same as [`vatio contacts`](https://vatio.ai/docs/cli/contacts); both are recorded in Activity |

Every change to the agent lands in **preview**, which only your team talks
to. Nothing reaches the people using your app until you publish, and
`rollback` undoes a publish. A workspace nobody outside your team has talked
to yet is a [draft](https://vatio.ai/docs/deployments#draft-and-published): there, live follows
preview, so every change reaches the app and its channels at once and there is
nothing to publish. The publish gate is the same as everywhere: when
there are reported replies, preview has to pass a test run first.

## Who can do what

The same roles as the rest of Vatio ([Team](https://vatio.ai/docs/team)):

- **The owner and developers** can do everything above.
- **Supervisors** can read conversations and report replies.
- **Viewers** can read conversations.

A workspace that requires two-step verification needs it on your account
before any tool works on it.

Reading a conversation from the MCP is recorded in the workspace's
**Activity**, like reading it in the inbox, and so is every change, with the
app that made it. `list_conversations` redacts emails, phone numbers and
other personal data; `get_conversation` shows a conversation as the developer
sees it, the way `vatio chat show` does.

## Your files and the MCP

If your agent was built with [vatio-init](https://vatio.ai/docs/integrations/claude-code), its
`vatio.yml` and `tools/` live in your app's repository, and `vatio push`
deploys them. An edit from the MCP changes the agent on Vatio and not your
files, so a later `vatio push` from the repository would undo it.

It does not. While preview carries edits your files have not seen,
`vatio push` is refused and lists them. `vatio pull` writes them into
`vatio.yml` and `tools/`, keeping your comments and everything else in the
file, and then you push as usual:

```bash
vatio pull
git diff
vatio push
```

[Edits made outside your files](https://vatio.ai/docs/cli/deploy#edits-made-outside-your-files)
has the details. `get_agent` says which edits preview carries that no push has
written over yet.