In 2025, teams spent the year wrapping every testing tool in an MCP server. In 2026, the question is sharper: do they really need the Model Context Protocol (MCP), or is the command-line interface (CLI) that their agent already knows good enough?
This session is a practical look at how a large language model (LLM) can drive an entire test stack - user interface (UI) testing with Playwright, application programming interface (API) checks, reporting in ReportPortal, accessibility scans, and issue tracking - as one coordinated pipeline instead of disconnected silos. The Model Context Protocol (MCP) is one way to expose those tools to the model with structured, governed access. But a fast-growing 2026 consensus says that, for most developer-facing tasks, letting the agent call plain CLIs is cheaper, more reliable, and closer to what the model was actually trained on. Benchmarks have shown CLI-driven agents running many times cheaper in terms of tokens and far more reliably than their MCP equivalents on identical tasks.

Michał Ślęzak will share real experiments from both worlds: where MCP genuinely earns its place - auth, governance, stateful tools, and anything without a CLI - where a CLI-first setup wins on speed and maintenance, and what a hybrid approach, now the production default for agents like Claude Code and Cursor, actually looks like in a test pipeline. Attendees will leave able to decide, tool by tool, when to reach for MCP and when Bash is simply better.