Prompt Optimizer MCP

The problem

Vague prompts waste tokens, produce weak outputs, and cost more than they should. Most people don't realize their prompt is ambiguous until after the LLM has already processed it. There's no pre-flight check, no way to know if your prompt will produce the output you actually wanted before spending the tokens.

What I built

An MCP server with 11 tools that acts as a prompt quality gate before execution. Works with any LLM: Claude, GPT, Gemini, or anything that supports MCP. The key constraint: zero AI calls inside. All analysis is deterministic: regex, heuristics, and rule engines. The optimizer itself costs nothing to run and adds negligible latency. Freemium model: 10 free optimizations, then Pro ($4.99/mo) or Power ($9.99/mo) tiers.

Core tools

• optimize_prompt: Analyzes raw prompts, detects ambiguities across 10 rules, scores quality 0-100, and compiles an optimized version
• refine_prompt: Iterative refinement loop that incorporates answers to blocking questions and re-scores
• approve_prompt: Sign-off gate that refuses to approve if blocking questions remain unanswered
• estimate_cost: Standalone token estimator for any text, with multi-provider pricing (Claude, GPT, Gemini)
• compress_context: Context pruner that strips irrelevant code and docs before submission (57% token savings in testing)
• check_prompt: Quick pass/fail check with score, top issues, and a suggestion. No session needed
• configure_optimizer: Set mode, strictness, thresholds, and default output target
• get_usage: Current usage count, limits, remaining quota, and tier info
• prompt_stats: Aggregated optimization stats: total count, average score, estimated savings
• set_license: Activate a Pro or Power license key (Ed25519 offline validation)
• license_status: Check current license tier and expiry

How it works

The optimizer classifies each prompt against 13 task-type patterns (code, writing, research, planning, analysis, communication, data, and more). It auto-detects audience, tone, and platform (across 19 audience patterns and 9 platforms including Slack, LinkedIn, and email), applies task-type-aware goal enrichment, and uses intent-first classification to prevent misclassifying writing prompts as code. It detects multi-task overload, identifies high-risk domains, and recommends the right model tier. The compiled output supports three targets: Claude XML, OpenAI system/user, and generic Markdown.

Why it matters

Every prompt that goes through this server is clearer, cheaper, and more likely to produce the output you actually wanted. The 57% context compression alone means you can fit substantially more relevant information into each API call. Because the entire system is deterministic, it runs at zero marginal cost: no API calls, no token spend, no latency. Free to start, with Pro and Power tiers for teams that need volume.