Case Study: How Cursor Wires MCP + Skills + Harness Together

Cursor is among the most successful AI coding products (ARR > $500M). The edge is not “best model”—competitors use the same GPT-4o and Claude. Harness engineering is the moat.

Product Timeline

2023.03  Cursor 0.1 — Tab completion (Copilot competitor)
2024.03  Cursor 0.30 — Chat + @codebase semantic search
2024.08  Cursor 0.40 — Composer multi-file edits
2025.01  Cursor 0.45 — Agent mode (autonomous loop)
2025.06  Cursor 0.50 — MCP + Background Agent
2026.01  Cursor 0.55 — Skills + enhanced Agent harness

Inflection: Chat → Agent—from AI-assisted editor to AI coding agent platform.

Architecture Layers

Layer 1: Codebase Index (Context Foundation)

On project open, Cursor pre-builds a semantic index:

source files → chunks → embeddings → vector index
                                         ↓
user question → semantic search → top-K files → agent context

Why Cursor beats raw API coding: it knows your repo.

@codebase = semantic search; @file = precise; @folder = directory scope. Context engineering as product.

Layer 2: MCP Integration (Tool Extension)

Mid-2025 MCP in Settings → MCP Servers:

Cursor Agent
  ├── built-ins: read_file, edit_file, terminal, search
  ├── MCP: GitHub (PRs, issues)
  ├── MCP: Postgres
  └── MCP: custom…

Example workflow:

User: "How many new signups last week? If abnormal, open a GitHub issue."

Agent:
  1. [MCP postgres] SELECT count(*) …
  2. [analyze] 40% below average
  3. [MCP github] create_issue(...)
  4. [reply] "Opened #123—234 signups, 40% below baseline"

Layer 3: Skills (Behavior Standardization)

2026 Skills in .cursor/skills/:

• load relevant skills by task type
• growing community (commits, PR review, SDK, etc.)
• Skills + Rules (.cursor/rules/) = project AI policy

Layer 4: Agent Harness (Loop Engine)

user goal
  → context manager (index + open files + skills + MCP)
  → LLM → tool calls or final answer
  → tool router (built-ins + MCP)
  → append results, loop (cap N, user can interrupt)

Design choices:

• human-in-the-loop (interrupt anytime)
• workspace sandbox
• terminal commands often need confirmation

vs Competitors

	Cursor	GitHub Copilot	Windsurf
Agent loop	✅ mature	⚠️ Workspace (newer)	✅ Cascade
MCP	✅	❌	✅
Skills	✅	❌ (instructions)	❌
Codebase index	✅ pre-built	⚠️ limited	✅
Human control	strong	medium	medium
Model choice	multi	OpenAI only	multi

Moat = index + MCP + Skills + loop—not the model weights alone.

Lessons for Builders

1. Context first—index ROI beats swapping models
2. Open tool interfaces—MCP grows capability via ecosystem
3. Standardize behavior—Skills capture team SOPs
4. Collaboration > full autonomy in coding UX
5. Eval-driven—tab accept rate, agent completion, undo rate

References

• Cursor MCP docs
• Cursor Agent mode
• Model Context Protocol

Summary

Cursor ≈ Codebase Index × MCP ecosystem × Skill standards × mature Agent harness. All four—not models alone.