MCP vs API vs CLI: how should an AI agent connect to tools?
Compare MCP, direct APIs and CLIs for AI agents by tool discovery, host support, credential storage and execution requirements.
By Scrollport
In brief
MCP standardises how an AI application discovers and invokes tools, an API exposes the underlying software operation, and a CLI gives an agent a local command interface. Use a CLI on a permitted persistent machine, a direct API from a programmatic harness with a secret store, and remote MCP when the AI host supports OAuth connectors but cannot safely retain a local credential.
MCP, APIs and CLIs solve different parts of an agent connection. MCP standardises how an AI application discovers and invokes tools, an API exposes the software operation, and a CLI gives an agent a local command interface. For Scrollport, use the CLI on a permitted persistent machine, the direct API from a programmatic harness with a secret store, and remote MCP when the AI host supports OAuth connectors but cannot safely retain a local credential.
MCP, APIs and CLIs are different layers
An API is the application interface behind an operation. A CLI packages commands for a shell and may call that API. MCP is a client-server protocol through which an AI host can discover and invoke exposed tools. A remote MCP server can still call APIs behind the scenes, while a CLI can expose the same product controls through local commands.
All three Scrollport routes reach the same five controls: apps, discover,inspect, run and read-only wallet. The catalog, account and wallet stay the same. The practical choice is where authority can be stored safely and which transport the surrounding host supports.
Compare MCP, API and CLI
| Connection | Best fit | Credential owner | Main trade-off |
|---|---|---|---|
| Remote MCP | An AI host with native remote connectors and no suitable persistent shell. | The host stores OAuth connector tokens. | Support and setup behaviour depend on the host. |
| Direct API | A hosted or embedded harness that already makes HTTPS requests. | The harness stores a durable bearer secret. | The team owns the client, storage and error handling. |
| CLI | An agent with a permitted shell and persistent writable storage. | The machine and operating-system user store the credential. | It is unsuitable for ephemeral or restricted environments. |
MCP vs API: protocol and execution layer
Choose remote MCP when the AI product already provides a connector interface and retains OAuth tokens for that connector. The host connects to Scrollport’s remote MCP server and exposes its tools to the model. The official MCP architecturedefines the client-server roles, and the MCP authorization specificationdefines OAuth-based authorization for HTTP transports.
Choose the direct API when a programmatic harness already has a durable secret store and can implement the device flow over HTTPS. The harness owns polling, credential retention, request handling and logs. It must keep the returned secret out of source code, logs and model context. If the runtime disappears after one request, it should not create a credential that will immediately become orphaned.
MCP vs CLI: discovery and local control
Choose the CLI when the agent has permission to run shell commands, write to persistent storage and reuse that storage in later sessions. The setup client starts browser authorization and stores the credential without asking the human to paste a raw secret into chat. The credential belongs to the machine and operating-system user, so name the connection by harness and access method in Agent access.
Remote MCP is better for connector-native hosts without that local environment. It is not automatically simpler: installation can require explicit human configuration and a new agent session. Its advantage is that the host already provides the discovery surface and a suitable place to retain connector authority.
A deterministic connection decision tree
- Reuse first. Test an existing Scrollport connection with free
walletordiscover. - Prefer CLI when the environment has a permitted shell and persistent writable storage.
- Use remote MCP when the host supports remote OAuth connectors but cannot safely keep a local credential.
- Use direct API when a programmatic harness can complete device authorization and retain the secret correctly.
- Stop if none of those environments can preserve authority safely.
Human approval remains separate in every route. Signing in identifies the human; it does not authorize the agent. Follow the setup guide for CLI and device flow, or the remote MCP guide for a connector-native host. If you are deciding whether to operate several MCP servers behind another layer, read the MCP gateway comparison next.