CLI vs MCP vs SDK

STACK has three client surfaces. They are not tiers of the same thing. Each one is built for a different driver doing a different job. Pick by who is at the wheel.

MCP server

Driver: an LLM assistant (Claude Code, Cursor, any MCP client). Use it for setup and exploration in natural language.

• Endpoint: mcp.getstack.run, over OAuth.
• Auto-discovered stack_* tools the assistant calls on your behalf.
• Best for: "register this agent and issue it a read-only Slack passport" without leaving your editor.

SDK

Driver: the agent's own code. Use it on the runtime hot path, in-process.

• Packages: @getstackrun/sdk (JS/TS), getstack (Python).
• Typed wrapper over every REST route, plus one thing the raw API cannot give you: offline passport verification (verifyPassportOffline) with no network call at verify time.
• Best for: an agent that verifies passports, issues sub-passports, or streams LLM calls through the gateway as part of its normal execution.

CLI

Driver: a human operator in a terminal. Use it for ops, scripts, and CI.

• Invocation: npx @getstackrun/cli <command>.
• A near-complete mirror of the operator surface: agents, passports, services, drop-offs, skills, identity, team, detectors, notifications, and more.
• Two things it is uniquely good at: a live audit-log tail (monitor --follow) and scriptable credential retrieval (service credential <provider>) for CI and .env-style workflows.
• Best for: an operator setting up or running an account from a shell, or a CI job that needs a scoped, audited, revocable credential.

Which one do I use?

• Setting things up by talking to an assistant → MCP.
• Writing agent code that runs at request time → SDK.
• Working from a terminal, a script, or CI → CLI.