One MCP server was 69% of an agent's starting context

Ioannou runs an internal company agent on OpenCode 1.18.31 (his deployment, not a benchmark). Token use was climbing faster than sessions or automations could explain. One conversation recorded 50.4 million tokens across 121 model calls, 49.9 million of them cache reads, which first suggested long conversations. Following sessions back to their first call showed an investigation starting at about 177K input tokens before it had gathered anything, and a fresh conversation answering "hi" showed 177,444.

The OpenCode TokenScope report put the skill catalog at about 3,400 tokens and the tool definitions it could inspect at about 5,800, which did not account for the request. It omitted the full MCP definitions. Inspecting those separately, he found that Meta Ads alone had 97 tools and roughly 595K characters of descriptions and parameter schemas, about 72% of the measured MCP definition text. With Meta Ads disabled, the same greeting dropped to 54,181 tokens. With it restored and OPENCODE_EXPERIMENTAL_CODE_MODE=true set, the greeting used 13,274, which he reports as about 92% smaller than the original.

He is explicit about the limit of that number: it is the greeting's recorded usage, "not a claim about our final bill." A real investigation still loads definitions and results as it works. He also notes that OpenCode 2 makes Code Mode the default for MCP servers, and that Claude Code's MCP Tool Search and Codex's deferred MCP discovery follow the same on-demand pattern. All figures are his own, from his hosted deployment, and he says an AI assistant helped reconstruct the investigation from session history.

Why it matters: If your agent has a long list of integrations, the cost of making every tool available on every request can dwarf the cost of using them, and a dashboard total will not show it. Check the first-call input size of a fresh session before blaming long conversations or reasoning effort.