Building Model Context Protocol (MCP) servers is less about adding capabilities and more about managing the hidden tax of context window pollution. In a recent audit of the nova-ops-nim-mcp server, developer eliashivemind evaluated three candidate tools for the server's third slot, ultimately deciding to keep the existing two-tool architecture intact. The decision wasn't driven by a lack of technical feasibility, but by a strict adherence to a token budget rule: keeping the total schema under 1,000 estimated tokens per turn and ensuring every property description remains a single short line.

The Cost of 'Just in Case' Features

The first candidate, nim_get_job_status, was priced at approximately 60 tokens per turn. While logically sound for asynchronous operations, the current v1.0.0 implementation of the generate calls is synchronous. Paying 60 tokens on every interaction for a tool whose precondition doesn't exist yet violates the core principle of avoiding 'just in case' registrations. The verdict was clear: this tool earns its slot the day the server goes async, not before.

Flexibility Often Means Bloat

The most tempting candidate was nim_create_media, a flexible tool designed to handle image generation, video creation, upscaling, and restyling via an operation enum. However, the audit revealed that this 'one tool to rule them all' approach cost 1,449 characters and roughly 363 tokens per turn. This is nearly double the cost of the entire existing server. The complexity arose because most parameters only applied to specific operations, forcing the schema to include conditional prose like 'required for generate_image... ignored otherwise.' This resulted in a bloated schema that charged rent for four tools while providing only one.

Decision Rules for Infrastructure Hygiene

The audit established a clear hierarchy for accepting new tools. First, a real workflow must already hurt without the feature; 'could use it' is insufficient justification. Second, the measured cost must fit within the budget with room to spare, anticipating future candidates. Finally, the one-line description rule serves as a diagnostic tool: if a parameter requires a paragraph to explain, the tool's shape is fundamentally flawed. The server remains at two tools and ~198 tokens per turn, proving that a budget's purpose is not to say no, but to make every 'yes' specific and earned.

Key Takeaways

  • Token budgets are hard constraints: All three candidates fit under the 1,000-token limit, yet two were rejected for not meeting workflow urgency or schema simplicity standards.
  • Synchronous vs. Asynchronous matters: Adding status polling tools to synchronous endpoints wastes context tokens on every turn.
  • Monolithic tools inflate schemas: Combining multiple operations into one tool with conditional parameters often costs more tokens than separate, focused tools.
  • One-line descriptions are a design check: If you can't describe a parameter in one line, your tool interface is likely too complex or poorly scoped.

The Bottom Line

Stop treating your MCP server's context window as infinite. If a tool doesn't solve an immediate, painful workflow bottleneck, it's just expensive noise in your prompt.