Opinion · Systems & Protocols
Polish speech recognition, high-order reasoning, and the Model Context Protocol are each commodities in 2026. No single product ships all three in one voice session.
There is a peculiar cruelty in the current generation of artificial intelligence, and it has nothing to do with runaway cognition. It is the cruelty of the near miss — a stack that executes almost every requirement and then fails the integration test on the last assertion, every single time.
The requirement is unremarkable. A user wants a voice interface he can drive hands-free, in Polish, backed by a frontier reasoning model, with the ability to invoke his own tools over the Model Context Protocol — MCP, the open standard that lets a model call external servers for reading files, writing files, querying private data. Three capabilities: multilingual automatic speech recognition, competent reasoning, and MCP tool-calling inside a spoken session. Each is a commodity. The paradox is that no shipping product composes all three.
Begin with ChatGPT. Its automatic speech recognition, powered by Whisper-class models, handles Polish flawlessly — input and output. But its MCP connectors are confined to the text surface. In Advanced Voice Mode, custom MCP servers are simply not exposed; the tool layer disappears the moment you switch from typing to speaking. Polish voice, yes. MCP over voice, no.
Now Claude. Here MCP is a first-class citizen of the conversation: you register a remote MCP server as a connector, tied to your subscription rather than a metered API token, and the model reasons over its tools — list_files, read_file, write_file — mid-dialogue. The defect is upstream, at the ASR layer: Polish is not an available voice-input language. It will read and write Polish text perfectly, but the microphone path demands English, which collapses the hands-free use case.
“Whisper transcribes Polish. The models reason. MCP is published and adopted. Every layer is a commodity — and the stack still does not exist.”
And Gemini completes neither axis. Its custom MCP support lives in Gemini Spark, gated behind an AI Ultra plan, a personal (non-Workspace) account, US-only availability, a mandatory Keep Activity setting — and, fatally, an English-only restriction. The broader consumer app relegates MCP to developer surfaces: the Gemini CLI and the enterprise console, over StreamableHTTP transport with a publicly reachable endpoint. The one house praised for multilingual voice cannot bind that voice to custom MCP tools.
What makes this maddening is abundance, not scarcity. Polish ASR is solved — open Whisper weights transcribe it faithfully, offline. Frontier reasoning is sold by the token. MCP is a ratified, widely adopted protocol with a growing server ecosystem. Every layer sits on the shelf. And still the user must become his own systems integrator: wire Whisper to a reasoning model via API, bolt on a text-to-speech engine for Polish output, mount his MCP servers, and orchestrate the loop — ASR to model to tool-call to result to TTS — by hand, merely to obtain what any reasonable person assumes already exists.
This is the commodity paradox of the current age. We were promised general intelligence. We received a market of brilliant specialists that will not compose. The frontier is not throttled by model capability. It is throttled by product boundaries — the deliberate fences each vendor draws around which layer meets which surface.
One day a single product will bind multilingual ASR, frontier reasoning, and subscription-tier MCP into one voice session without a weekend of glue code. Until then, the state of the art remains a magnificent set of components — and the whole, that stubborn, ordinary whole, stays just out of reach.