I Benchmarked 10 MCP Servers — One Burns 47,000 Tokens Just to Say Hello
benchmarks
| Source: Mastodon | Original article
A benchmark of ten popular MCP servers found one server burns 47 K tokens just to say hello, using more tokens than an entire GPT‑3 conversation.
A new independent benchmark of ten widely‑used MCP (Model‑Centric Programming) servers has revealed a startling inefficiency: one implementation consumes roughly 47 000 tokens merely to return a greeting. The test, conducted over three weeks, catalogued 847 tools across the servers and parsed 312 000 tokens of JSON schema data, exposing a token‑bloat problem that dwarfs the cost of a typical GPT‑3 conversation.
The findings echo earlier observations from the community. A Medium post by Andy Nguyen noted that some MCP servers can burn 20 000 tokens per call, while a Frontend Architect’s June 2026 note warned that connecting ten servers may expend more than 75 000 tokens before a user types anything. OnlyCLI’s “MCP Token Trap” article explained that agents automatically inject the full tool catalog into the LLM context, turning a 93‑tool GitHub server into a 55 000‑token payload. The new benchmark confirms that this token overhead is not an edge case but a systemic issue across popular implementations.
Why it matters is twofold. First, token consumption translates directly into higher API costs for developers and enterprises that rely on LLM‑backed tooling. Second, the inflated context window leaves less room for actual user prompts, degrading the responsiveness and relevance of AI‑assisted workflows. In a landscape where efficiency is a competitive edge—highlighted in our recent coverage of AI‑agent credit usage—such waste threatens to erode the economic viability of LLM‑driven development pipelines.
Looking ahead, the community is likely to scrutinise the JSON‑schema handling and tool‑catalog injection mechanisms that drive the excess. Upcoming releases of MCP servers may adopt more compact schema representations or lazy‑loading strategies, as suggested by performance work on language‑specific JSON libraries (e.g., Go’s sonic versus stdlib). Watch for updates from the maintainers of the most token‑hungry servers and for broader discussions on standardising token‑efficient MCP protocols.
Sources
Back to AIPULSEN