Developers Converge: MCP Now Beats CLI for Most Agent Integrations

Between Sept 15-17, a wave of developer statements shifted the Skill+CLI vs MCP server debate firmly toward MCP: Shopify CEO Tobi Lütke's hands-on conclusion became the landmark voice, echoed by a Claude Code engineer and Nous Research co-founder Teknium, while mitsuhiko supplied a key diagnosis of why codemode underdelivers.

Confirmed

• On Sept 15, PrinceSauromates opened the discussion: with open-source models getting more reliable, he wanted to upgrade his AI coding workflow and personally leaned toward Skill+CLI for stronger control, but hesitated facing MCP's standardized integration.
• On Sept 16, Tobi Lütke said that, unexpectedly, MCP is now better than CLI for most integrations, citing greatly improved tool calling, deferred tools, and MCP now being stateless; trq212 relayed and endorsed the conclusion, adding that composing or filtering data can be handled by adding parameters like query to MCP tools. itsOmSarraf and other repliers agreed.
• On Sept 17, a Claude Code engineer shared the same counterintuitive judgment: improved tool calling makes structured MCP tools more reliable than wrapped command lines, and MCP now supports a stateless mode (relayed by seantstrong).
• Teknium replied to trq212 in agreement: at least for remote MCP, MCP now beats CLI integration.
• trq212 further reported that codemode-style implementations (having the model write code to invoke tools) did not work as well as expected, with reasons still under reflection.
• mitsuhiko diagnosed that codemode underperforms mainly because existing MCP servers rarely return stable, structured tool outputs; he found OpenAPI-style APIs pair better with codemode. He also agreed MCP now suits most integrations better than CLI.
• A view relayed by HankYeomans resonated widely: frontier models handle MCP comfortably, tools can be lazily loaded, and MCP beats exposing raw CLI commands in most scenarios.
• Rhys Sullivan (relayed by teropa) declared "total victory" for MCP: indexable tool catalogs let agents scale to unlimited tools, no full sandbox needed, and consistent auth across MCP servers.
• yenkel resurfaced his article "If you kill MCP, you don't give a st about security," arguing that CLI-to-API calls cannot provide the guarantees good security requires. In discussion with trq212 he drew finer boundaries: for reliable tool calling alone, bash is no longer a universal answer because it is harder to secure (halting problem among others), but bash + sandbox still fits code generation and execution work.

Unconfirmed

• PrinceSauromates's original intuition that CLI offers stronger control was neither verified nor refuted; trq212 said the reasons codemode underdelivers are "still being thought through," and mitsuhiko's structured-output diagnosis has not yet become consensus.

Why it matters

• The debate directly shapes infrastructure choices at the AI agent tool layer: if MCP's indexable catalogs, deferred loading, and statelessness are truly mature, CLI integration's weaknesses in scalability and auth consistency get amplified.
• With a Claude Code team member and Teknium joining in, the judgment has risen from individual developer experience to an engineering consensus among frontline agent teams and the model community.
• mitsuhiko's diagnosis reveals the other half of the problem: codemode works only if tools return stable, structured outputs, raising the bar for MCP server implementation quality.
• yenkel's follow-up adds nuance: plain bash is no longer a universal answer, but bash + sandbox retains a place for code generation and execution tasks, and the security community does not accept the "CLI replaces MCP" narrative — the debate is not fully closed.

2026-09-15 ~ 2026-09-17 · 14 related posts

• Skill+CLI vs MCP: a developer asks which integration route wins on tokens and reliability — PrinceSauromates · 2026-09-15
• MCPs Beat CLIs for Most Integrations, Developer Concludes After Hands-On Testing — trq212 · 2026-09-16
• [source] MCP now beats CLIs for most integrations as protocol goes stateless — mitsuhiko · 2026-09-16
• MCP security researcher: killing MCP means you don't care about security, CLIs lack guarantees — yenkel · 2026-09-16
• MCP's total victory: why MCP beats CLIs for most agent integrations — teropa · 2026-09-16
• MCP beats CLI for most integrations as models get better at tool calling, say AI engineers — HankYeomans · 2026-09-16
• [source] Shopify CEO Tobi Lütke: MCP now beats CLIs for most integrations — tobi · 2026-09-16
• Dev finds codemode implementations less useful than expected, fueling MCP vs CLI debate — trq212 · 2026-09-16
• mitsuhiko: unstable tool outputs explain codemode failures; OpenAPI + embeddings work great — mitsuhiko · 2026-09-16
• Claude Code devs: better tool calling makes MCP beat CLIs for most integrations — sean_t_strong · 2026-09-17
• [source] Bash is no longer all you need for reliable agent tool calling — yenkel · 2026-09-17
• Teknium agrees: with better tool calling, MCP now beats CLIs for most integrations — Teknium · 2026-09-17
• Devs argue MCP now beats CLIs for most integrations as models get better at tool calling — itsOmSarraf_ · 2026-09-17

1 near-duplicate retellings: trq212