> ## Documentation Index
> Fetch the complete documentation index at: https://ifdocdemo.lumivescence.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Connect an AI agent

> Give Claude Code, Cursor, Codex, or an MCP client access to InFlow through skills, plugins, or the MCP server.

An AI agent can run the InFlow CLI for you. The CLI gives the agent discovery, enrollment, and payments. Connect the agent in one of four ways.

## Ask the agent to set itself up

For the fastest setup, give your agent this instruction.

```text theme={null}
Fetch inflowcli.ai/skill.md and get me set up with InFlow.
```

The agent installs the CLI, starts the login, and shows you the approval URL. The CLI prints the same setup guide.

```bash theme={null}
inflow --bootstrap
```

## Install the skills

The CLI has three skills. Each skill is a playbook for one task.

| Skill | Task |
| - | - |
| `agentic-discovery` | Find services, offerings, and actions with ODP. |
| `agentic-enrollment` | Enroll with an AEP service and manage its credentials. |
| `agentic-payments` | Pay MPP and x402 endpoints. |

Install all three for an agent that supports skills.

```bash theme={null}
npx skills add inflowpayai/inflow-cli
```

An agent can also load one skill from the installed CLI. This copy always matches the installed version.

```bash theme={null}
inflow --skill agentic-payments
```

The web copies are at `inflowcli.ai/skills/<name>.md`. They follow the latest release. Compare their `version:` header with `inflow --version`.

Load the playbook before the task. Setup alone does not teach the agent the safe order of commands.

## Run the CLI as an MCP server

The CLI can run as an MCP server that exposes its commands as tools.

```bash theme={null}
inflow --mcp
```

Add a server named `inflow` to your MCP client configuration.

```json theme={null}
{
  "mcpServers": {
    "inflow": {
      "command": "inflow",
      "args": ["--mcp"]
    }
  }
}
```

If your MCP client cannot find `inflow`, set `command` to the full path. To get the full path, run `command -v inflow`.

The tool names replace the spaces in a command with underscores.

Do not use `npx @inflowpayai/inflow --mcp`. The npm package does not start MCP. Install the CLI first. Then use `inflow --mcp`.

Some MCP hosts do not load skills. For these hosts, put the output of `inflow --skill <name>` into the system prompt.

## Install a plugin

The [CLI repository](https://github.com/inflowpayai/inflow-cli) includes plugins for Claude Code, Cursor, and Codex. In Claude Code, run these commands.

```text theme={null}
/plugin marketplace add inflowpayai/inflow-cli
/plugin install inflow@inflow
```

## Let the agent read the CLI

An agent can read the current commands from the installed CLI.

| Command | Output |
| - | - |
| `inflow --llms` | The command index. |
| `inflow --llms-full` | The full manifest with every parameter. |
| `inflow <command> --schema` | The JSON Schema for one command. |

To set the output format, see `--format` in the [CLI reference](/reference/cli#global-options).

## What stays with the owner

Three steps always need the owner.

| Step | Why |
| - | - |
| Unlock the vault | The CLI reads the PIN or the passphrase only from a terminal. |
| Approve the login | The owner approves the device flow in a browser. |
| Approve payments | InFlow sends a payment request to the owner when approval is necessary. |

## Supported agents

| Type | Names |
| - | - |
| Coding agents and assistants | Claude Code, ChatGPT, Codex, Cursor |
| Agent frameworks | Manus, OpenClaw, NemoClaw, Hermes, and custom MCP clients |

## Next step

<Card title="Pay for a resource" icon="credit-card" href="/agents/pay-for-a-resource">
  Read how the agent pays, and what the owner approves.
</Card>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.