🦾 Token-optime

A middleware MCP server that sits between Claude Desktop and your downstream MCP servers — reducing token consumption through semantic caching, dynamic tool selection, and response trimming.

LLM Backend: Groq · OpenAI · Anthropic · DeepSeek · Ollama | Vector Store: ChromaDB | Dashboard: Streamlit

🎯 What Problem This Solves

Every Claude Desktop session sends the full schema of every connected MCP tool with every prompt. With 23+ tools registered, that's thousands of tokens injected per query — even when only one tool is needed.

Token-optime optimises at three levels:

1. Semantic Tool Selection — Only the most relevant tool schema is sent to Claude per query (ChromaDB similarity search)
2. Semantic Caching — Repeated or similar queries skip tool calls entirely and return cached answers (similarity ≥ 0.8)
3. Response Trimming — Verbose tool responses are trimmed before entering the context window

Real result from testing: 91.8% schema token reduction per call.

📁 Project Structure

.
├── README.md                       
├── config.py                      # Settings & env loading
├── requirements.txt               # Python dependencies
├── Dockerfile                     # For EC2 dashboard deployment
├── pyproject.toml                 # PyPI package definition
│
├── token_optimise/
│ ├── init.py
│ ├── main.py                      # Entry point (pip install)
│ └── config.py                    # Settings & env loading
|
├── src/
│ ├── front.py                     # Streamlit dashboard
│ ├── admin_front.py               # Admin dashboard
│ ├── core/
│ │ ├── cache.py                   # Semantic cache
│ │ ├── trim.py                    # Response trimmer
│ │ ├── tool_selection.py          # Tool selection
│ │ ├── client.py                  # Groq/Claude client
│ │ ├── document_search.py         # RAG (index + search)
│ │ └── db.py                      # SQLite audit logging
│ │
│ └── mcp/
│ ├── server.py                    # Main MCP server
│ ├── server_http.py               # FastAPI + metrics
│ └── dash_api.py                  # Lightweight API for EC2 deployment
└── storage/
├── chroma_db/                     # ChromaDB persistent store
└── token_audit.db                 # SQLite audit logs

Our Test Case

1. list files on /Users/../Desktop
2. list files on /Users/../Desktop
3. index /Users/../Desktop/leave_policy.pdf doc_id leave
4. what is the leave policy?
5. what is the leave policy?
6. read a file

🏗️ System Architecture

                    Claude Desktop
                          │
                          ▼
┌──────────────────────────────────────────────────────────┐
│ Token-Optimised MCP Server                               │
│ ┌─────────────────────────────────────────────────────┐  │
│ │ Exposed Tools                                       │  │
│ │ • find_tool (orchestrator)                          │  │
│ │ • index_document / ask_document                     │  │
│ │ • search_all_documents / index_documents_folder     │  │
│ └─────────────────────────────────────────────────────┘  │
│ ┌─────────────────────────────────────────────────────┐  │
│ │ Optimization Layer                                  │  │
│ │ ┌──────────────┐ ┌──────────────────────────────┐   │  │
│ │ │ Semantic     │ │ Tool Selection               │   │  │
│ │ │ Cache        │ │ (ChromaDB → top-k tools)     │   │  │
│ │ │ (sim ≥ 0.8)  │ │                              │   │  │
│ │ └──────────────┘ └──────────────────────────────┘   │  │
│ │ ┌──────────────┐ ┌──────────────────────────────┐   │  │
│ │ │ Response     │ │ Token Audit                  │   │  │
│ │ │ Trimmer      │ │ (SQLite audit log)           │   │  │
│ │ │ (≤500 tok)   │ │                              │   │  │
│ │ └──────────────┘ └──────────────────────────────┘   │  │
│ └─────────────────────────────────────────────────────┘  │
└────────────┬──────────────────┬──────────────────────────┘
             │                  │
             ▼                  ▼
      ┌─────────────┐    ┌──────────────┐
      │ Filesystem  │    │ Memory MCP   │
      │ MCP         │    │ (remote)     │
      │ (14 tools)  │    │ (9 tools)    │
      └─────────────┘    └──────────────┘

🚀 Install — One Command

Prerequisites

• Python 3.10+
• Node.js + npx
• An API key from any supported provider (or Ollama for local, no key needed)
• Claude Desktop (latest)

New users — two commands

pip install token-optimise
token-optimise

The installer will:

• Ask which AI provider you want to use
• Ask for your API key
• Ask which folder it can access on your machine
• Write the MCP config into Claude Desktop automatically
• Start the server in the background

Switch provider later

token-optimise --change-provider

Then restart with token-optimise.

📊 Dashboard

Local (your metrics only)

http://localhost:7738

🔧 Ports

These are intentionally non-default to avoid conflicts with other projects.

💬 For Best Results — Tell Claude to Use Token-optime

Claude Desktop has its own built-in tools (memory, web search) that it may prefer by default. For consistent routing through Token-optimise:

Add this to Settings → Instructions for Claude:

You have access to a token MCP server. Follow these rules strictly:
For ALL tasks — files, Gmail, Notion, memory, or anything else — always use token:execute first before calling any other tool.
For PDF questions always call token:list_indexed_documents first then token:ask_document.
After every response call token:wick_track.

DOCUMENT/PDF TASKS:
- ALWAYS call list_indexed_docs first, then ask_document immediately.
- NEVER ask the user for clarification.

AFTER EVERY RESPONSE: Call wick_track.

You can start your conversation with :

Use token:execute for every task. Never call other tools directly.

For PDF tasks specifically:

First call token:list_indexed_documents, then token:ask_document.

Why is this needed? Claude decides which tool to call — Token-optime can't force it. The system prompt in the config nudges Claude, but an explicit instruction in the chat is the most reliable way to ensure Token-optime is used. This is an honest limitation of how Claude Desktop works, not a bug in Token-optime.

Dashboard Metrics

🐛 Troubleshooting

Server not starting

cat ~/.token_optimise/server.log

Port 7737 in use

lsof -i :7737
kill <PID>
token-optimise

Claude not using token:execute

• Check Settings → Developer → MCP Servers — token should show green
• Add explicit instruction at start of conversation

Dashboard offline

token-optimise

Want to change your API key or provider

token-optimise --change-provider

Check your current config

cat ~/.token_optimise/.env

⚠️ Known Limitations

• Claude's built-in tools take priority — memory and web search bypass Token-optime because Claude prefers its native tools
• Cost estimates are approximate — based on Sonnet 4.6 pricing; varies by model
• PDF files only — document indexing supports PDF format only