Thariq flips, says MCP now beats CLIs for most integrations
The "just give it a CLI" consensus is getting walked back by several of the people who held it.
Thariq posted the reversal plainly. "I was not expecting things to go this way, but I think MCPs are better than CLIs for most integrations," he wrote, giving three reasons. Models have gotten much better at tool calling, tools can be deferred, and MCP is now stateless. His practical note for anyone building one is that if you need to compose or filter data, add params like query to your MCP tools.
I was not expecting things to go this way, but I think MCPs are better than CLIs for most integrations.
Armin Ronacher agreed, with a boundary on it. MCP today is a massive improvement over where it started, he said, "at least on the protocol level". That is a narrower endorsement than Thariq's, and worth reading as one.
The cost argument moved
dax made the same shift about economics rather than ergonomics. "now that we have codemode you can add as many MCPs as you want without cost," he wrote, though he says he still does not add a ton up front and instead picks them up as he hits use cases. His claim is cumulative. Over time his agent gets more and more capable, especially with project specific servers.
now that we have codemode you can add as many MCPs as you want without cost
Peter Steinberger was blunter about whether the debate is still live. Replying to @zeeg, he said "that argument lost all meaning with code mode".
that argument lost all meaning with code mode
Wes Bos used the moment to send people back to something they skipped. If you wrote off WebMCP because all you needed was a CLI, revisit that, he said, calling WebMCP key. Kent C. Dodds, for his part, claimed the call, saying he was expecting things to go this way and inviting followers to come for more good bets.
For working engineers, the thing that changed is not taste. The old objection to MCP was that every server you registered burned context whether or not the agent used it, which made a plain CLI the cheaper way to hand an agent a capability. The reversals here all point at the same fix, tool definitions no longer have to sit in context up front. Nobody in this thread posted numbers, so treat it as informed practitioner judgment rather than a measured result. If you stripped MCP out of your setup on the old advice, it is worth retesting.