SemaCache
Cut your LLM bill by caching answers you have already paid for. Exact and semantic matching in front of OpenAI, Anthropic, Gemini and xAI, with an agent-readable control panel for hit rate, spend and tuning.
- Access to all tools via gateway
- 5,000 calls/month
- Cancel anytime
Secure payment via Stripe. Cancel anytime.
SemaCache sits between your application and your LLM provider and stops you paying twice for the same answer.
Every product built on an LLM asks the same questions over and over. The same document gets summarised, the same support question gets asked in fifty slightly different ways, the same prompt gets retried after a timeout. You pay full price every time. SemaCache remembers, and the second time it is free.
It matches two ways. Exact, for identical prompts - the easy case. And semantic, for prompts that mean the same thing in different words, which is the case a hash-based cache always misses and the one that actually happens in production. "Summarise this contract" and "give me a summary of this contract" are one answer, not two.
Connecting is two lines: point your OpenAI-compatible client at the proxy URL and use your SemaCache key. Nothing else in your code changes. You keep using your own OpenAI, Anthropic, Gemini or xAI account, so your billing relationship with the provider is untouched - the only difference is that repeated work stops reaching them. One endpoint covers all four providers; the right one is picked from the model name.
What this listing adds on top is a control panel your agent can read. Ask how the cache is doing and get hit rate, estimated savings and latency percentiles back as data rather than a screenshot. Ask why the hit rate is low and get the recent queries, each marked hit or miss - near-duplicates showing as misses is the signature of a threshold set too strictly, and the agent can move it and measure the difference in the same conversation.
The traffic path stays out of MCP on purpose. Routing inference through a tool call would mean either pasting a provider key into an agent transcript or having the marketplace fund your inference. Instead your application talks to the proxy directly and the tools observe it - which also means the cache keeps working at 3am when no agent is involved.
Full product at www.semacache.io.
Use cases
Find out how much caching is actually saving you, in dollars
Diagnose a low hit rate and fix the matching threshold in the same conversation
See which model is eating the budget before the invoice tells you
Catch an expired provider key the moment requests start failing
Prove a prompt change worked by comparing hit rate before and after
Cut latency on repeated prompts from seconds to milliseconds
Features
Drop-in OpenAI-compatible proxy - change your base URL and API key, nothing else
Exact matching on identical prompts, semantic matching on paraphrases a hash cache always misses
Works across OpenAI, Anthropic, Gemini and xAI from one endpoint - the provider is detected from the model name
Text, image and video responses all cached
Estimated savings in dollars, not just a hit-rate percentage
Per-model breakdown of volume, spend and latency, plus p50/p95/p99
Recent queries marked hit or miss, so a badly-set threshold is visible rather than guessed at
Tune similarity threshold and TTL from the agent, then measure the effect
Your provider key is encrypted at rest and never passed as a tool argument
2M managed embedding tokens and 25,000 stored vectors included
Frequently asked questions
Q.What is SemaCache, in one sentence?
A caching proxy that sits between your application and your LLM provider, so the second time someone asks the same question - or a differently-worded version of it - you get the answer back instantly and pay nothing for it. See www.semacache.io.
Q.How do I connect my app?
Two lines. Point your OpenAI-compatible client at the proxy base URL and use the SemaCache API key instead of your provider key. Both appear on your Subscriptions page as soon as you save your provider key. Your own key stays server-side at SemaCache and is never sent to your agent.
Q.Do I need a SemaCache account?
No. Your mcpmarket subscription is the only account you need. Calls are authenticated with a short-lived token issued per request, so your marketplace key is never shared with the upstream service.
Q.Whose provider key pays for the LLM calls?
Yours. SemaCache caches your own OpenAI, Anthropic, Gemini or xAI account - it does not resell inference. Your billing relationship with the provider is unchanged; the only difference is that repeated work stops reaching them. The $9 covers the cache and the control panel — the same price as SemaCache Pro bought directly.
Q.Where does my provider key go?
You enter it once on your Subscriptions page. It goes straight to SemaCache, encrypted at rest, and is never stored by mcpmarket. It is deliberately not settable through a tool call: anything passed as a tool argument is written into your agent's transcript.
Q.What is semantic matching, and when does it help?
Exact matching only fires on byte-identical prompts. Semantic matching embeds the prompt and reuses an answer when a previous one means the same thing - so "summarise this contract" and "give me a summary of this contract" share one answer. That is where most of the savings come from in a real product, because users never phrase things identically.
Q.What does lowering the similarity threshold do?
It reuses responses for prompts that are merely similar rather than near-identical, which raises the hit rate and lowers spend, at some risk of answering a subtly different question. 0.90 to 0.98 is the useful band. Change it, then check get_cache_stats and get_recent_cache_queries to see what it did.
Q.Does the cache slow anything down?
A hit returns in milliseconds without touching the provider. A miss adds one lookup before passing through. Cache-Control: no-store on a request skips the cache entirely when you need a guaranteed fresh answer.
Q.Is my data shared with anyone?
No. Every cache entry, vector and setting is scoped to your subscriber id, and the tools only ever read your own traffic.
Reviews
No reviews yet. Be the first to share your experience.
Quick start
- 1
Subscribe using the button above
- 2
Create an API key in Dashboard
- 3
Add the MCP to Cursor, Claude, or your client
Gateway URL
https://www.mcpchannel.ai/api/mcpOne URL for every MCP you subscribe to. Authenticate with your API key in an X-MCP-Market-Key header.