Llmforge MCP — Data skill for Claude Code

LLMForge MCP Server 🧠⚡🔬

Universal Autonomous AI Engineering: LLM Architecture, Evals, Fine-Tuning, Inference Cost Optimization & Production LLMOps Engine MCP Server for **Cursor, Claude Code, Windsurf, and Autonomous AI Agents**.

🌟 Core Capabilities:

• 👑 Master 1-Shot Outcome Engine (llm.architect): Ingests the AI feature concept to synthesize a complete LLM blueprint in one call: approach decision matrix (prompting vs RAG vs fine-tuning vs agents — and when each is wrong), model selection matrix (frontier flagships vs small tiers vs open-weights), token cost projections at scale with cascade math, gateway/routing/caching architecture, and a prototype-to-production roadmap.
• 🔬 Eval Harness Builder (llm.evals): Golden dataset design (sizes, sourcing, contamination hygiene), task-matched strategies (classification exact-match, RAG triple-metrics, agent trajectory grading, code execution pass@1), LLM-as-judge with bias controls (binary rubrics, judge-human calibration), and CI regression gates that block quality-dropping deploys — with complete runnable harness code.
• 🧬 Fine-Tuning Recipe Engine (llm.finetune): Honest should-you-even-finetune gates (knowledge injection → RAG, not training), base model selection by compute budget (free Colab to A100 clusters), dataset preparation that wins or loses the run, complete LoRA/QLoRA configs with concrete hyperparameters (rank, alpha, LR, epochs), tuned-vs-base evaluation protocols, and deployment paths (vLLM multi-LoRA, GGUF edge).
• 💸 Inference Cost Slasher (llm.inference_costs): Prioritized savings plans with expected percentages — prompt caching (30-60% of input costs), model cascade routing with confidence escalation (40-70%), Batch API (flat 50%), output token discipline, semantic caching, distillation endgames, and self-hosting break-even math.
• 🛡️ Production Hardening Kit (llm.production): Layered guardrails (prompt injection defense-in-depth, schema enforcement, PII scru