Hidden Cost Black Holes in MCP Tool Calls
C00LDude6ix9ine · reddit · 2026-07-15
While troubleshooting cost overruns across multiple agent systems, the author discovered that calculating costs per entire run reveals almost no waste; the real focus should be on the cost of individual MCP tool calls.
The post lists several common issues:
- The same tool being called repeatedly, often due to a lack of session-level caching
- Retrieval / memory looking like a "memory cost" when it actually includes massive embedding and retrieval overhead never used by the final answer
- Tool schemas and long descriptions being repeatedly counted as input tokens on every call
- Many observability dashboards only showing total agent costs, making it impossible to pinpoint "which tool is burning money"
The author offers a rough calculation: if a single tool call costs $0.003 and runs an extra 10,000 times daily, that one tool could add roughly $900 in monthly expenses.
More from coding & agent
- Codex helps build Valdiluce, an open-world game with climbing, gliding and gondolas — Dimillian · 2026-07-22
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- LangSmith adds tracing for Pipecat, LiveKit, OpenAI Realtime, and Gemini Live — LangChain · 2026-07-22
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Annotated transcript of a Claude Code team interview is now available — trq212 · 2026-07-22