Skip to content

Repository files navigation

The Best Copilot Setup Ever Built

What Is This?

This is a complete, portable toolkit for building the most powerful GitHub Copilot development environment that exists. It turns VS Code + Copilot Pro+ from a code completion tool into an autonomous multi-agent engineering system with persistent memory, self-improvement, and workflow pipelines.

No one — in any documented open-source project, community repo, blog post, or expert configuration — has built anything like this. We checked. Exhaustively. 66KB of beast-mode research across official docs, GitHub, npm, Stack Overflow, Medium, archives, and obscure sources confirmed: this is novel.

The Problems This Solves

Problem 1: Copilot Forgets Everything

Every Copilot session starts from zero. You explain your project, your conventions, your tech stack, your preferences — and next session, it's all gone. Context compaction (when the conversation gets too long) wipes it again mid-session.

Our solution: A persistent brain system with 7 memory files that survive sessions, compaction, and context loss. A compaction survival anchor (CLAUDE.md) that the AI reads first after any reset. Structured handoff protocols that transfer state between sessions like shift notes in a hospital.

Problem 2: Copilot Is a Generalist

Default Copilot gives the same quality response whether you're asking it to debug a race condition, design a database schema, or audit for SQL injection. It has no specialization.

Our solution: 9 specialist agents, each with domain expertise, curated tool access, and model routing. The security auditor uses Opus (deepest reasoning, 3x cost) because security demands it. The scout uses GPT-5 mini (free) because file lookups don't need expensive models. Each agent knows its job, has its tools, and hands off to the next agent in the pipeline.

Problem 3: Copilot Doesn't Learn

If something fails, Copilot doesn't remember. It'll try the same failed approach again next session. It doesn't track what works, what doesn't, or what decisions were made and why.

Our solution: A self-improvement protocol with dead-end tracking, lessons-learned accumulation, and a self-grading system that evaluates every major task on 6 criteria. The system literally grades itself and fixes gaps before reporting "done."

Problem 4: No Workflow Orchestration

Default Copilot handles one task at a time. There's no concept of "research, then design, then build, then review, then security audit, then performance test."

Our solution: 7 chat modes with handoff buttons that create workflow pipelines. Feature pipeline: Planning → Architect → Build → Review → Security → Performance. Fix pipeline: Scout → Debugger → Build → Review. Each stage gates the next — if security finds a critical issue, you fix it before proceeding.

Problem 5: Cost Waste

Copilot Pro+ gives you ~1500 premium requests/month with model multipliers. Using Claude Opus (3x) for simple questions burns your budget. Using GPT-5 mini (free) for architecture decisions wastes your time.

Our solution: Model routing baked into every agent and chat mode. 60% of interactions are free (GPT-5 mini), 25% are standard (Sonnet 4.6, 1x), only 5% use heavy reasoning (Opus, 3x). Budget usage: ~650 of 1500 requests/month with massive headroom. Plus an automatic escalation rule: if the cheap model struggles twice, it tells you to switch up.

Problem 6: No External Intelligence

Default Copilot can only see your code. It can't search the web, read current documentation, query your database, run browser tests, or manage version control programmatically.

Our solution: 10 MCP (Model Context Protocol) servers that give Copilot super-powers:

  • Tavily — Deep web search, page extraction, site crawling
  • Context7 — Up-to-date library documentation (not stale training data)
  • Playwright — Browser automation and visual testing
  • PostgreSQL — Direct database queries from chat
  • Sequential Thinking — Deep reasoning chains for complex problems
  • GitHub — Full repo management (issues, PRs, branches, code search)
  • Memory — Knowledge graph for persistent entity relationships
  • Filesystem — Direct file operations
  • Git — Full version control (diff, log, blame, branch)
  • Codebase Memory — Tree-sitter indexed knowledge graph (64 languages, sub-ms queries)

What You Get

When this toolkit is fully deployed, your Copilot environment includes:

Component Count Purpose
Specialist Agents 9 Domain experts (architect, security, DB, DevOps, etc.)
Chat Modes 7 Pre-configured contexts with model routing
Reusable Skills 7 Multi-step capabilities (research, audit, pipeline, etc.)
Prompt Templates 9 One-click workflows (beast mode, debug, review, etc.)
Scoped Instructions 9 Language/framework-specific coding standards
MCP Servers 10 External intelligence (web, docs, DB, git, browser, etc.)
Brain Memory Files 7 Persistent knowledge that survives sessions
VS Code Extensions 8+ Developer toolchain (linting, Docker, testing, etc.)
Workflow Pipelines 4 Feature, Fix, Research, Quality pipelines
Self-Grading 6 criteria Automatic quality evaluation after every task
Auto Hooks 3 Syntax validation on file save

How to Use This Toolkit

Option A: Automated Setup (Recommended)

  1. Copy this entire BestCoPilotSetup/ folder into your project root
  2. Open VS Code with Copilot Pro+
  3. Start a new chat session (preferably in Build mode or with Claude Sonnet 4.6+)
  4. Paste this prompt:
    Read the file BestCoPilotSetup/GO.md and execute every instruction in it. 
    Ask me the startup questions first, then build everything autonomously.
    
  5. Answer the 4 setup questions when prompted
  6. Let the AI build everything — it will create 50+ files, configure settings, install extensions, and set up MCP servers
  7. When it's done, it will run beast mode research to find anything new that could improve the setup

Option B: Manual Setup

  1. Read PREREQUISITES.md — install required software
  2. Read ARCHITECTURE.md — understand every component
  3. Use configs/ — copy settings, MCP configs, extension lists
  4. Use framework/ — adapt and copy .github files
  5. Set up brain templates from brain/
  6. Run beast mode research from BEAST-MODE-RESEARCH.md

File Guide

File What It Is
README.md You're reading it — the master overview
GO.md THE key file — the AI bootstrap prompt that builds everything
PREREQUISITES.md Software, accounts, and keys you need before starting
ARCHITECTURE.md Deep technical explanation of every component
FRAMEWORK-REFERENCE.md Complete content of every framework file (the "source code")
configs/vs-code-settings.jsonc All VS Code settings to add/merge
configs/mcp-servers.jsonc All 10 MCP server configurations
configs/extensions.md Extensions to install
brain/README.md How the brain/memory system works
brain/templates/ Empty templates for all 7 brain files
BEAST-MODE-RESEARCH.md The final research directive to keep improving

Requirements

  • VS Code 1.100+ (latest stable recommended)
  • GitHub Copilot Pro+ subscription ($39/month — gives access to all models, 1500 premium requests)
  • Node.js 20+ (for MCP servers that use npx)
  • Git installed and configured
  • PostgreSQL (optional — only if your project needs a database)
  • API keys: Tavily (free tier available), GitHub PAT (for GitHub MCP)

See PREREQUISITES.md for the complete list.


Philosophy

This system was built on a few core beliefs:

  1. AI should be a partner, not an assistant. It should think, decide, and act — not present menus and wait for permission.
  2. Memory is everything. A system that forgets is a system that stagnates. Persistent memory turns sessions from isolated events into a continuous learning curve.
  3. Specialization beats generalization. A security auditor agent that thinks like an attacker will find vulnerabilities that a generalist never would.
  4. Cost efficiency matters. Using the right model for the right task isn't penny-pinching — it's engineering discipline.
  5. Self-improvement is non-negotiable. The system grades itself, logs failures, and adapts. Stagnation is failure.

Built across 5 sessions of intensive development. March 2026.

About

Portable toolkit for building the ultimate VS Code Copilot agent framework — 9 agents, 7 chat modes, 10 MCP servers, persistent brain system. Paste GO.md into Copilot and it builds everything autonomously.

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors