> **Download Free CLAUDE.md ($0)**: https://pulse-forge.io/api/launch-kit/free-claude-md?repo=getcursor/cursor | **1-Click Drop-In Config Pack ($4.99)**: https://pulse-forge.io/buy/micro_pass?repo=getcursor/cursor&source=md_stack | **Complete 9-File BOM .ZIP ($9.99)**: https://pulse-forge.io/buy/repo_pass_999?repo=getcursor/cursor&source=md_stack_999

# ADR-2026-AI-STACK: Production AI Architecture Decision Record for `getcursor/cursor`
- **Status**: PROPOSED (Curated via PulseForge 12-Slot Deterministic BOM Engine)
- **Team & Workload Profile**: `25` Engineers | `500,000` Monthly Agent Sessions | Mode: `SOC2_COMMERCIAL`
- **Configured Stack Slots**: `10` Active Layers | `2` Intentionally Omitted Layers (Lean YAGNI Setup)
- **Total Monthly TCO**: `$3,260.00/mo` (vs. `$6,351.80/mo` Unoptimized Baseline -> **Saves `$3,091.80/mo`**)
- **Concurrent Local RAM Footprint:** `2175 MB` | **Cumulative p50 Overhead:** `71.8 ms` | **Min Local VRAM:** `0.0 GB`

## 1. Curated Active Bill of Materials (BOM) & License Audit
| Slot # | Architectural Layer | Selected Tool (`owner/repo`) | License | Cloud Egress | p50 Overhead | RAM (50 Agents) | Est. Monthly Cost |
|---:|---|---|---|---|---:|---:|---:|
| 1 | **AI Coding IDE & Editor** | `cursor/cursor` (Cursor Pro) | `Proprietary` | Cloud Proxy | `18.0 ms` | `640 MB` | `$1,100.00/mo` |
| 2 | **Autonomous Terminal & SWE Agent** | `anthropics/claude-code` (Claude Code CLI) | `Proprietary` | Cloud Proxy | `22.0 ms` | `190 MB` | `$1,200.00/mo` |
| 3 | **Agent Harness & Context OS** | `affaan-m/ECC` (ECC v2.2 (Everything Claude Code Harness)) | `MIT` | Zero-Egress Local/Self-Host | `1.8 ms` | `45 MB` | `$75.00/mo` |
| 4 | **LLM Gateway, KV-Cache & FinOps Router** | `Portkey-AI/gateway` (Portkey AI Gateway) | `MIT` | Zero-Egress Local/Self-Host | `1.9 ms` | `115 MB` | `$190.00/mo` |
| 5 | **Multi-Agent Orchestration & State Graph** | `pydantic/pydantic-ai` (PydanticAI Type-Safe Agents) | `MIT` | Zero-Egress Local/Self-Host | `2.4 ms` | `220 MB` | `$155.00/mo` |
| 6 | **Structured Output & Schema Compiler** | `BoundaryML/baml` (BAML (Rust Schema-Aligned Parser)) | `Apache-2.0` | Zero-Egress Local/Self-Host | `0.9 ms` | `55 MB` | `$90.00/mo` |
| 7 | **Agent Memory, Vector & Hybrid RAG Engine** | `qdrant/qdrant` (Qdrant Quantized Vector Engine) | `Apache-2.0` | Zero-Egress Local/Self-Host | `3.4 ms` | `420 MB` | `$145.00/mo` |
| 8 | **Code Sandbox, MicroVM & Browser Runtime** | `daytonaio/daytona` (Daytona Self-Hosted Agent Workspace) | `Apache-2.0` | Zero-Egress Local/Self-Host | `18.0 ms` | `310 MB` | `$180.00/mo` |
| 9 | **MCP Server Runtime & Tool-Poisoning Firewall** | `invariantlabs-ai/mcp-scan` (Invariant MCP-Scan Firewall) | `Apache-2.0` | Zero-Egress Local/Self-Host | `1.4 ms` | `60 MB` | `$50.00/mo` |
| 10 | **50-Probe OWASP Red-Team & Regression CI Gate** | `promptfoo/promptfoo` (Promptfoo 50-Probe OWASP Red-Team) | `MIT` | Zero-Egress Local/Self-Host | `2.0 ms` | `120 MB` | `$75.00/mo` |

## 1B. Intentionally Omitted Layers (Lean YAGNI Justification)
- **Slot 11 (Voice, WebRTC & Realtime Multimodal Agent)**: `OMITTED BY DESIGN` - Text/code-only workload; WebRTC voice & realtime audio pipeline not required.
- **Slot 12 (Local Inference & QLoRA / GRPO Fine-Tune Engine)**: `OMITTED BY DESIGN` - Using routed frontier API models; dedicated local GPU/VRAM inference & QLoRA fine-tuning not required.

## 3. Recommended Drop-In Arbitrage Swaps
- **AI Coding IDE & Editor**: Swap `cursor/cursor` -> **`cline/cline`** to save **`$775.00/mo`** & **`430 MB RAM`** (Eliminate $20/seat markup and cloud indexing egress via BYOK Cline).
- **Autonomous Terminal & SWE Agent**: Swap `anthropics/claude-code` -> **`Aider-AI/aider`** to save **`$775.00/mo`** & **`65 MB RAM`** (Pair with Aider tree-sitter repo-map to cut discovery tokens by 58%).