cap-shield

Context selection for AI agents. Measured, not estimated.

pip install cap-shield

from cap_shield import CapShield

cap = CapShield(api_key="cap_live_...")

cap.remember("support", "Customer reports a delayed parcel, order 4471")
ctx = cap.context("support", "what did the customer complain about?")

print(ctx.text)              # send this to your model
print(ctx.saving_pct, "%")   # how much you did not send
print(ctx.left_behind)       # ...and how many entries were left behind

What this does

Your agent's history grows every turn, and the whole thing is billed on every call. This selects what answers the question and leaves the rest.

Measured on our corpus with cl100k_base: 80.5 % of context not sent at a 1000-token budget. Measured on LongMemEval-S, 500 questions: 93.8 % retrieval recall against 51.9 % for word matching.

That second number is the one that matters. A saving is worthless if what was dropped is what your agent needed — so we publish both.

What this does not do

It does not call your model. ctx.text goes to Claude, GPT or whatever you use. This library never sees your model provider's keys.

It is not a proxy. Your uptime does not depend on ours for the model call itself.

Measure before you commit

No account needed:

CapShield(api_key="anything").measure(
    texts=["...your actual messages..."],
    query="what your agent would search for")

Runs on your own traffic and stores nothing.

Two mechanisms, deliberately separate

context() saves tokens — what your model is billed for. pack() saves bytes — bandwidth and storage. The packet is decompressed before the model sees it, so it saves no tokens.

Conflating them is the most common misunderstanding about this product.

Full documentation

https://cap-shield-robin.fly.dev/docs/quickstart