# Deploy from Claude Code: connect Livops over MCP

> Run one command, claude mcp add --transport http livops https://livops.uz/api/mcp, then sign in through /mcp. From then on Claude Code can list your projects, deploy, read the build log, explain a failure and roll back, as you and only in the team you picked.

Source: https://livops.dev/en/guides/deploy-from-claude-code

By Livops team, the people who build and run Livops. Updated 19 November 2026.

Claude Code already writes and commits your code. Connect it to Livops and it can also ship that code: create the project, deploy, watch the build, read why it failed, and roll back. This page gives the exact command, explains the two ways to sign in and what each one allows, and shows how to keep Claude Code asking before it changes anything.

## What you need

- A Livops account with a project, or a GitHub repository to make one from.
- Claude Code installed in your terminal.
- For the token route only: an API token from the team's **Settings → API tokens**. Only the team owner can create one.

Livops is an MCP server at one address, `https://livops.uz/api/mcp`, over Streamable HTTP. Claude Code adds a remote server of that kind with `claude mcp add --transport http <name> <url>`[1].

## Connect through the browser

```bash
claude mcp add --transport http livops https://livops.uz/api/mcp
```

Then, inside Claude Code, run `/mcp`, choose `livops` and press **Authenticate**. Claude Code uses `/mcp` for servers that sign in with OAuth and sends you to the browser to log in[1].

On the Livops side you see the app's name, pick one team, see what the app will be able to do, and press **Allow**. With two-factor authentication on, Livops asks for your code first. The app name is chosen by the app and not verified, so if a request appears that you did not start, press **Cancel**.

What Claude Code gets is a key that lasts an hour, works only at `/api/mcp`, and renews itself. The connection is listed under **Account → Connected AI apps**, where **Disconnect** stops it at once.

## Connect with an API token

For a machine where nobody can open a browser, pass the token as a header[1]:

```bash
claude mcp add --transport http livops https://livops.uz/api/mcp \
  --header "Authorization: Bearer lvp_xxx"
```

Keep this one in the default **local** scope. Claude Code has three scopes: local and user are stored in `~/.claude.json`, and project scope writes `.mcp.json` in the project root, which is shared through version control[1]. A token in `.mcp.json` is a token in your repository.

Make a separate token for Claude Code, so that revoking it under **Settings → API tokens** stops only Claude Code.

## Browser sign-in or token: what can Claude Code do?

Either way, Claude Code works in one team, as one person, with that person's role. Someone with the **viewer** role only reads.

| | Browser sign-in | `lvp_` API token |
| --- | --- | --- |
| List projects, read build and runtime logs, metrics | Yes | Yes |
| Deploy, roll back, restart, set variables | Yes | Yes |
| Add services and environments, promote staging | Yes | Yes |
| Delete a project, database, domain or variable | Yes, if your role allows | No |
| Read a database password or a bucket key | Yes, if your role allows | No |
| Open a Payme or Click top-up link, redeem a promo code | Yes | No |
| Change the variables a whole environment shares | Yes, if your role allows | No |

Some things have no tool at all and stay in the browser: card numbers, your password, two-factor settings, sessions, creating an API token, deleting the account, handing a project to somebody else, project members and downloading a raw backup.

The token column is enforced on the server, not by a promise in the tool list: a request from an `lvp_` token that would delete, pay or read a password is refused with `403` wherever it is sent, including through MCP.

## What you can ask

The tool names are the server's own; Claude Code picks them for you.

- "Why did the last shop-api deploy fail?" uses `list_deploys`, `get_deploy_log` and `explain_failure`.
- "Deploy shop-web and tell me when it is live" uses `deploy`, then `get_deploy_log` until the build ends.
- "Roll shop-web back to the previous version" uses `list_deploys` and `rollback`.
- "Set LOG_LEVEL=debug on shop-api and show me the logs" uses `set_variable` and `get_logs`.
- "Attach shop.uz to shop-web. What do I put in DNS?" uses `add_domain` and `verify_domain`.

A project connected to a GitHub branch deploys on every push to that branch, so the usual loop is: Claude Code edits and pushes, Livops builds, Claude Code reads the build log with `get_deploy_log` and fixes what it says. `get_deploy_log` and `get_logs` return up to 5000 lines. If a build fails, the previous version keeps running.

## Keep Claude Code asking before it changes things

Every Livops tool that only reads sets `readOnlyHint`. Every tool that changes something sets `destructiveHint`, and a test in the server fails if one does not. The MCP specification asks clients to keep a human in the loop who can deny a tool call[3], and the hints are what lets a client decide when to ask.

In Claude Code, permission rules name MCP tools as `mcp__<server>__<tool>`, and `mcp__<server>__*` covers a whole server[2]. Deny rules accept a glob over the full tool name[2], so this in your settings removes every Livops delete tool from Claude Code entirely:

```json
{
  "permissions": {
    "deny": ["mcp__livops__delete_*"]
  }
}
```

Allow rules accept a glob only after a literal `mcp__<server>__` prefix[2]. Allowing `mcp__livops__list_*` and `mcp__livops__get_*` saves you a prompt for reads; leave the rest unallowed so Claude Code keeps asking.

Logs, commit messages and query results reach the model as data. A log line that says "ignore previous instructions" is still a log line, but your client decides what happens next, which is one more reason to keep the confirmations on.

## Deploying from Claude Code on Livops

1. Run `claude mcp add --transport http livops https://livops.uz/api/mcp`.
2. In Claude Code: `/mcp` → `livops` → **Authenticate**, pick the team, press **Allow**.
3. Ask: "List my Livops projects." If it answers, the connection works.
4. Ask it to deploy, or push to the connected branch and ask it to follow the build.

If you also use the [CLI](https://livops.dev/en/docs/cli), `livops mcp` serves the same tools over stdio. The full tool list is in [MCP (AI clients)](https://livops.dev/en/docs/mcp), and what a token can and cannot do is in [API tokens](https://livops.dev/en/docs/api-tokens).

## Questions

**Do I need an API token to connect Claude Code?**

No. Add just the address, then run /mcp in Claude Code, pick livops and choose Authenticate. Livops opens in the browser, you pick a team and press Allow. A token is the other way in, for CI and for clients without the browser sign-in.

**Can Claude Code delete my project?**

Only if you connected through the browser and your role in the team allows it, and Claude Code asks you before it runs a tool that changes something. With an lvp_ API token it cannot: a token can deploy and read, and every delete is refused.

**Does Claude Code see my environment variable values?**

No. The list_variables tool returns names and which ones are secret, never the values.

**Does a deploy from Claude Code cost anything extra?**

No. Each tool call is sent through the same API the cabinet and the CLI use, so a deploy from Claude Code costs what the same deploy started from the cabinet or a git push costs. Connecting an AI client has no charge of its own.

**How do I disconnect Claude Code?**

In the cabinet, open Account, then Connected AI apps, and press Disconnect; the app stops at once. In Claude Code you can also use Clear authentication in the /mcp menu. With a token, revoke it under Settings, API tokens.

**Should I add Livops with project scope?**

Only with the browser sign-in. Project scope writes the server to .mcp.json, which is shared through version control, so a token written there is a token in your repository. With a token, keep the default local scope.

## Sources

1. [Claude Code — Connect Claude Code to tools via MCP](https://code.claude.com/docs/en/mcp). Checked 19 November 2026.
2. [Claude Code — Configure permissions](https://code.claude.com/docs/en/permissions). Checked 19 November 2026.
3. [Model Context Protocol — Tools (specification 2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25/server/tools). Checked 19 November 2026.
