# Manifest gateway now has a CLI and a remote MCP server

> Source: <https://manifest.build/blog/manifest-cli-and-mcp/>
> Published: 2026-09-13 00:00:00+00:00

# Manifest gateway now has a CLI and a remote MCP server

The Manifest gateway is now drivable two new ways: a management CLI, **`mnfst`**, and a remote **MCP server**. Both call the same management API the dashboard does, so you can create harnesses, connect providers, set routing, and read the request log from your terminal or from an AI client.

## Manage the Manifest gateway from the terminal

Log in once, in the browser:

```
mnfst login
```

`mnfst login` opens the dashboard, you approve, and the CLI gets a token. The token never travels through the browser.

From there, the whole workspace is a command away:

```
mnfst agent create --name coding-assistant --platform openclaw --category coding
mnfst provider connect xai --auth-type api_key --credential-env XAI_API_KEY
mnfst agent configure coding-assistant --models grok-4.5,grok-4 --provider xai
mnfst routing test coding-assistant
mnfst requests get --agent coding-assistant --range 7d
```

The command set covers harnesses, provider connections, routing (the default chain, custom tiers, [Autofix](/docs/autofix/), and recording), models and prices, the request ledger, and a `doctor` command that tells a wrong URL apart from a wrong credential in one pass.

`mnfst run` is the one built for coding agents. It injects a harness key into a child process without putting it in `argv` or the transcript:

```
mnfst run --agent coding-assistant -- python agent.py
```

The CLI is part of the Manifest repository today:

```
git clone https://github.com/mnfst/manifest
cd manifest && npm install
npm run build --workspace=packages/cli
npm link --workspace=packages/cli
```

## Connect any MCP client

The Manifest gateway also runs a remote [Model Context Protocol](https://modelcontextprotocol.io) server, so MCP clients can manage the workspace directly. It speaks MCP over HTTP and signs in with OAuth: your client opens a consent screen the first time, you approve, and it keeps a short-lived, revocable token.

Point any MCP client at:

```
https://app.manifest.build/api/v1/mcp
```

On a self-hosted install, use your own origin: `https://your-host/api/v1/mcp`.

### Claude Code

```
claude mcp add --transport http manifest https://app.manifest.build/api/v1/mcp
```

### Codex

```
codex mcp add manifest --url https://app.manifest.build/api/v1/mcp
codex mcp login manifest
```

### OpenCode

Add the server to `opencode.json`:

```
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "manifest": {
      "type": "remote",
      "url": "https://app.manifest.build/api/v1/mcp",
      "enabled": true
    }
  }
}
```

Then authorize it:

```
opencode mcp auth manifest
```

Any other MCP client works the same way: add the URL and approve in the browser.

The consent screen asks for read-only or read-and-write access. A read-only connection can look but not touch, and the write tools are not exposed to it at all. The tools mirror the CLI, including a real `routing test` that sends one request through the harness route to prove it works.

The full setup, discovery endpoints, and tool list are in the [MCP server docs](https://manifest.build/docs/integrations/mcp/).

## Get started

- Dashboard: [app.manifest.build](https://app.manifest.build)
- CLI: [packages/cli](https://github.com/mnfst/manifest/tree/main/packages/cli)
- MCP server docs: [manifest.build/docs/integrations/mcp](https://manifest.build/docs/integrations/mcp/)
- Source: [github.com/mnfst/manifest](https://github.com/mnfst/manifest)
