Token Optimizer Mcp

ooples/token-optimizer-mcp

This is a token reduction layer that sits between Claude and your file system, API calls, and development tools. It replaces standard MCP tools like read_file and grep with cached, compressed alternatives that store content in SQLite and return only diffs on subsequent calls. The 80+ tools cover file operations, API fetching with TTL-based caching, database queries, test runners, and multi-tier cache management with LRU/LFU eviction. Real production data shows 60-90% token reduction across operations. You'd reach for this when you're hitting context limits on large codebases or repeatedly fetching the same API responses, and you want automatic optimization without changing your workflow.

GitHub stars: 578
tools: 7
transport: STDIO
env var: 1
last updated: Nov 11, 2026

View on GitHub

Tools

7 tools

Public tool metadata for what this MCP can expose to an agent.

proxy_chatSend a chat completion request through PromptThin's cost-saving proxy. Five optimization routes are applied automatically: (A) semantic cache, (B) prompt compression via LLMLingua 2, (C) model routing to cheaper models, (D) context pruning at 8K tokens, (E) thinking budget cap...4 params

Send a chat completion request through PromptThin's cost-saving proxy. Five optimization routes are applied automatically: (A) semantic cache, (B) prompt compression via LLMLingua 2, (C) model routing to cheaper models, (D) context pruning at 8K tokens, (E) thinking budget cap...

Parameters* required

max_tokensinteger

Maximum tokens to generate. Higher values allow longer responses but cost more.default: 1000

messagesarray

Conversation messages in OpenAI chat format. Each message needs a 'role' and 'content'.

modelstring

The LLM model to use. Examples: 'gpt-4o', 'gpt-4o-mini' (OpenAI), 'claude-sonnet-4-6' (Anthropic), 'gemini-2.5-flash' (Gemini), 'llama-3.1-8b-instant' (Groq).

temperaturenumber

Controls randomness (0.0 = deterministic, 2.0 = very random). Lower for factual tasks, higher for creative.default: 0.7

proxy_predictEstimate cost savings BEFORE making a real LLM call — completely free, no tokens consumed. Returns original token count, estimated tokens after savings, cost comparison, saving percentage, applicable methods, and a recommendation.3 params

Estimate cost savings BEFORE making a real LLM call — completely free, no tokens consumed. Returns original token count, estimated tokens after savings, cost comparison, saving percentage, applicable methods, and a recommendation.

Parameters* required

messagesarray

Conversation messages to analyze for savings estimation. Same format as chat_completion.

modelstring

The target LLM model to estimate costs for. Examples: 'gpt-4o', 'claude-sonnet-4-6', 'gemini-2.5-flash'.

providerstring

The LLM provider. Must match the model — e.g. 'openai' for gpt-4o, 'anthropic' for claude models.one of openai · anthropic · gemini · groq

billing_start_trialStart a 7-day free Pro trial. Returns a Stripe checkout URL. No charge for 7 days. Pro plan: $4.99 first month, then $11.99/month. 10,000 requests/month. Cancel anytime.1 params

Start a 7-day free Pro trial. Returns a Stripe checkout URL. No charge for 7 days. Pro plan: $4.99 first month, then $11.99/month. 10,000 requests/month. Cancel anytime.

Parameters* required

formatstring

Response format: 'text' for a human-readable message with the checkout URL, 'json' for machine-parseable output.one of text · jsondefault: text

usage_summaryRetrieve complete usage summary: total requests, cache hit rate, total tokens saved, estimated cost saved in USD, and requests routed to cheaper models.1 params

Retrieve complete usage summary: total requests, cache hit rate, total tokens saved, estimated cost saved in USD, and requests routed to cheaper models.

Parameters* required

formatstring

Response format: 'text' for a human-readable summary, 'json' for structured data.one of text · jsondefault: text

billing_statusRetrieve current plan, monthly request limit, requests used this month, remaining requests, and usage percentage. Warns when usage exceeds 80%.1 params

Retrieve current plan, monthly request limit, requests used this month, remaining requests, and usage percentage. Warns when usage exceeds 80%.

Parameters* required

formatstring

Response format: 'text' for a human-readable status, 'json' for structured data.one of text · jsondefault: text

cache_flushMark all cached responses as stale, forcing fresh LLM calls. Use after updating a knowledge base or changing system prompts.1 params

Mark all cached responses as stale, forcing fresh LLM calls. Use after updating a knowledge base or changing system prompts.

Parameters* required

confirmboolean

Safety flag to confirm the flush operation. Must be true to proceed. Prevents accidental cache invalidation.default: true

usage_recentRetrieve recent proxied API requests with timestamp, provider, model, token counts, cache hit status, routing info, and tokens saved.1 params

Retrieve recent proxied API requests with timestamp, provider, model, token counts, cache hit status, routing info, and tokens saved.

Parameters* required

limitinteger

Number of recent requests to return. Defaults to 10. Maximum is 50.default: 10

Configuration

TOKEN_OPTIMIZER_CACHE_DIR

Directory path for cache storage (optional, defaults to ~/.token-optimizer-cache)

Readme

Token Optimizer