# CAP-Shield

Context selection and compression for AI agents.

Two things, measured separately and never added together:

- **Compression** saves BYTES over the wire. Packets are decompressed
  before a model sees them, so this does NOT reduce inference cost.
- **Memory retrieval** saves TOKENS in the context, by selecting the
  relevant entries instead of sending the whole history.

## Try it without an account

Two of the five MCP tools need no key. You can measure the saving on
your own traffic before trusting anything here.

- MCP over HTTP: `https://cap-shield-robin.fly.dev/mcp`
- Without MCP: `POST https://cap-shield-robin.fly.dev/api/v1/try`
- Python: `pip install cap-shield`

## Everything measured

All figures — including the packages that measure BADLY, the share of
messages that came out LARGER, and what has NOT been measured — are
published machine-readably and generated from running code:

https://cap-shield-robin.fly.dev/.well-known/cap-shield.json

**No numbers are repeated in this file.** They would go stale the moment
the dictionaries are retrained, and nothing here would notice. Fetch the
document instead.

## For agents

- Tool descriptions and when NOT to use them: https://cap-shield-robin.fly.dev/SKILL.md
- MCP server source, single file, no dependencies: https://cap-shield-robin.fly.dev/cap_mcp.py
- Standalone unpacker, no account needed: https://cap-shield-robin.fly.dev/cap_unpack.py
- Fixed benchmark report against /api/v1/try, no account needed: https://cap-shield-robin.fly.dev/cap_bench_report.py
- Human-readable quickstart: https://cap-shield-robin.fly.dev/docs/quickstart

Registered in the official MCP registry as
`io.github.robinlidberg-dot/cap-shield`.