Token Saver MCP — AI as a Full-Stack Developer

Transform AI from a code suggester into a true full-stack developer — with instant access to code intelligence and real browser control.

📚 Full Usage Guide & Examples → |
📖 Detailed Technical README → |
🔄 Releases

🚀 What is Token Saver MCP?

Modern AI coding assistants waste enormous context (and your money) by stuffing full grep/search results into the model window. That leads to:

❌ Slow lookups (seconds instead of milliseconds)
❌ Thousands of wasted tokens per query
❌ AI “losing its train of thought” in cluttered context

Token Saver MCP fixes this.

It gives AI assistants direct access to VSCode’s Language Server Protocol (LSP) and the Chrome DevTools Protocol (CDP), so they can work like real developers:

Instantly navigate & refactor code
Run code in a real browser (Edge/Chrome)
Test, debug, and verify changes themselves

Result: 90–99% fewer tokens, 100–1000× faster responses, and $200+ in monthly savings — while enabling AI to truly act as a full-stack engineer.

✨ Why Token Saver?

Think of your AI’s context window like a workbench. If it’s cluttered with logs, search dumps, and irrelevant snippets, the AI can’t focus.

Token Saver MCP keeps the workbench clean.

🔍 Without Token Saver

grep -r "renderProfileImage" .
# 5000+ tokens, 10–30 seconds, bloated context

⚡ With Token Saver

get_definition('src/components/UserCard.js', 25)
# 50 tokens, <100ms, exact location + type info

Cleaner context = a sharper, more persistent AI assistant.

🏗️ Revolutionary Dual Architecture

Token Saver MCP uses a split architecture designed for speed and stability:

AI Assistant ←→ MCP Server ←→ VSCode Gateway ←→ VSCode Internals
               (hot reload)   (stable interface)

🏗️ VSCode Gateway Extension
- Installed once, rarely updated
- Exposes VSCode’s LSP via HTTP (port 9600)
🚀 Standalone MCP Server
- Hot reloadable — no VSCode restarts
- Language-agnostic (JS/TS, Python, Go, Rust…)
- Bridges MCP protocol ↔ VSCode Gateway + CDP (port 9700 by default)

Why it matters: You can iterate on MCP tools instantly without rebuilding/restarting VSCode. Development is 60× faster and much more reliable.

🧰 What You Get

Token Saver MCP currently provides 40 production-ready tools across five categories:

LSP Tools (14) → get_definition, get_references, rename_symbol, get_hover, find_implementations, …
Memory Tools (9) → smart_resume (86-99% token savings vs /resume), write_memory, read_memory, search_memories (full-text search), export_memories, import_memories, …
Browser Tools (8) → navigate_browser, execute_in_browser, take_screenshot, get_browser_console, …
Testing Helpers (5) → test_react_component, test_api_endpoint, check_page_performance, …
System Tools (4) → get_instructions, retrieve_buffer, get_supported_languages, …

📚 See the full Usage Guide with JSON examples →

📊 Proven Results

Operation	Traditional Method	With Token Saver MCP	Improvement
Find function definition	5–10s, 5k tokens	10ms, 50 tokens	100× faster
Find all usages	10–30s	50ms	200× faster
Rename symbol project-wide	Minutes	100ms	1000× faster
Resume context (/resume)	5000+ tokens	200-500 tokens	86-99% savings

Token & Cost Savings (GPT-4 pricing):

Tokens per search: 5,000 → 50
Cost per search: $0.15 → $0.0015
Typical dev workflow: $200+ saved per month

🌐 Browser Control (Edge-Optimized)

Beyond backend code, Token Saver MCP empowers AI to control a real browser through CDP:

Launch Edge/Chrome automatically
Click, type, navigate, capture screenshots
Run frontend tests & debug JS errors in real-time
Analyze performance metrics

Example workflow:

AI writes backend API (LSP tools)
AI launches browser & tests API (CDP tools)
AI sees error logs instantly
AI fixes backend code (LSP tools)
AI verifies fix in browser

➡️ No more “please test this manually” — AI tests itself.

🧠 Smart Memory System (NEW!)

Replace wasteful /resume commands with intelligent context restoration:

The Problem with /resume

Dumps entire conversation history (5000+ tokens)
Includes irrelevant tangents and discussions
Costs $0.15+ per resume
AI gets lost in the noise

The Solution: Smart Resume

smart_resume()  // 200-500 tokens, focused context only

Features:

86-99% token savings compared to /resume
Progressive disclosure: Start minimal, expand as needed
Full-text search: Find memories by content, not just keys
Importance levels (1-5): Critical info persists, trivia can be dropped
Verbosity levels (1-4): Control detail granularity
Time-based filtering: Resume work from specific periods
Export/Import: Backup and share memory contexts between sessions

Example:

// Standard resume - just the essentials
smart_resume()  

// Include everything from last 3 days
smart_resume({ daysAgo: 3, verbosity: 3 })

// Critical items only for quick check-in
smart_resume({ minImportance: 4, verbosity: 1 })

Memory is stored locally in SQLite (~/.token-saver-mcp/memory.db) with automatic initialization.

🖥️ Real-Time Dashboard

Visit http://127.0.0.1:9700/dashboard to monitor:

Server status & connection health
Request metrics & response times
Token & cost savings accumulating live
Tool usage statistics

Perfect for seeing your AI’s efficiency gains in action.

⚡ Quickstart (30 Seconds)

# Clone repo
git clone https://github.com/jerry426/token-saver-mcp
cd token-saver-mcp

# One-step setup
./mcp setup /path/to/your/project

That’s it! The installer:

Finds open ports
Creates config files
Tests connection
Provides the Claude/Gemini command

➡️ Full installation & build steps: Detailed README →

🔌 Supported AI Assistants

Claude Code → works out of the box with MCP endpoint
Gemini CLI → use /mcp-gemini endpoint
Other AI tools → MCP JSON-RPC, streaming, or simple REST endpoints available

Endpoints include:

http://127.0.0.1:9700/mcp (standard MCP)
http://127.0.0.1:9700/mcp-gemini (Gemini)
http://127.0.0.1:9700/mcp/simple (REST testing)
http://127.0.0.1:9700/dashboard (metrics UI)

🔬 Verify It Yourself

Think the claims are too good to be true? Run the built-in test suite:

python3 test/test_mcp_tools.py

Expected output shows: hover, completions, definitions, references, diagnostics, semantic tokens, buffer management, etc. — all passing ✅

🛠️ Development

pnpm install
pnpm run dev       # hot reload
pnpm run build
pnpm run test

MCP server lives in /mcp-server/, with modular tools organized by category (lsp/, cdp/, helper/, system/).

See Full Technical README → for architecture diagrams, tool JSON schemas, buffer system details, and contributing guide.

📍 Roadmap / Vision

Token Saver MCP already unlocks full-stack AI workflows. Next up:

🔧 More browser automation tools (multi-tab, network control)
📦 Plugin ecosystem for custom toolpacks
🌐 Multi-assistant coordination (Claude + Gemini + others)
🧠 Expanded context management strategies

📄 License

MIT — free for personal and commercial use.

👉 Start today:

Run ./mcp setup
Tell your AI: “Use the get_instructions tool to understand Token Saver MCP.”
Watch your AI become a focused, cost-efficient, full-stack developer.

📚 For in-depth details:

Usage Guide & Examples
Full Technical README

Token节省器