# tare MCP server

> Reduces LLM context window usage by 44–73% through lossless, cache-safe, and output-aware compression for coding agents.

- Repository: https://github.com/mstuart/tare
- Page: https://hostedmcp.org/servers/mstuart/tare
- Transport: —
- Stars: 5
- License: MIT

## Environment variables

| Name | Required | Secret | Description |
|---|---|---|---|
| `TARE_UPSTREAM` | no | no |  |

## FAQ

### How does tare handle context compression?

It uses lossless-by-default compression that re-encodes tool outputs, logs, and JSON into denser equivalent forms, while remaining cache-correct for provider prefix caches.

### Can I use tare with my Claude subscription without an API key?

Yes, by running tare-mcp as a local stdio server, your agent calls it as tools without the server needing to call the model itself, thus requiring no API key.

### Does tare break provider prefix caching?

No, it is designed to be cache-correct by detecting the cache breakpoint and only compressing the dynamic suffix.
