MCP vs API vs CLI: how AI agents should connect to your app
rseroter · x · 2026-11-06
Wasp published a 46-minute deep dive comparing three ways to make a web app agent-friendly:
- MCP shines in the agent-native world because it tells agents what they should do, benefiting both server and client (e.g. Claude Code, ChatGPT desktop), and suits human-in-the-loop tasks.
- APIs and CLIs emphasize what agents can do, mostly server-side.
Advice for builders:
- Build a high-quality MCP server and REST API on one layer of operations: put server logic in plain functions and make each surface a thin adapter.
- Invest most care in the MCP server — the best agent experience, on a spec that keeps improving.
- Use the API for code that runs on its own: raw exports, bulk writes, nightly syncs, third-party integrations, and scripts agents rerun (many MCP clients lack code mode).
- A CLI is probably unnecessary unless agents must read local files — but it's cheap to expose over a good API.
Commenter rseroter notes the piece doesn't discuss turn/token costs per approach and may undersell the CLI.
Related event: Wasp Compares MCP vs API vs CLI for AI Agent Integration(4 posts)→
More from coding & agent
- Engineer's portfolio spans AI agents, a NASA sleeping pod and an oil painting — theteknosaur · 2026-11-06
- Adding 80+ biomes to Minecraft: AI-generated specs for ores and textures — Promptmethus · 2026-11-06
- Write decision logs: five lines per decision can save your team a whole meeting — dSebastien · 2026-11-06
- Gemini CLI PR fixes OOM in long agentic sessions with 64KB tool output budget and inlineData eviction — tadipavan · 2026-11-06
- Technical explainer: how the open-source Row-Bot agent handles memory — Acceptable-Object390 · 2026-11-06
- Hygieia: open-source ontology audit tool for AI app health — n0hatr3d0nlyhe4rt · 2026-11-06