Julius cuts token spend.

A token saver for your coding agent. Verbose commands typically cost 60 to 90% fewer tokens, with the same signal. Measured, recoverable and never lossy where it matters.

View on GitHub (opens in a new tab)Get started free
install with Homebrew
$ brew install hoophq/tap/julius
# then hook it into Claude Code, in every project:
$ julius init -g
$ julius doctor

See it in action.

Pick a command and run it. Julius drops the noise and keeps failures, branch state and hunks verbatim. If filtering would not shrink the output, you get the original.

without julius6 lines · 113 chars
$ go test -v ./...
=== RUN   TestA
--- PASS: TestA (0.01s)
=== RUN   TestB
--- PASS: TestB (0.00s)
PASS
ok  	github.com/x/pkg	0.212s
what reaches the model
$ julius go test -v ./...
Run the demo to filter this output before it reaches the model.

Each case is a test from Julius’s built-in filter pack, input and expected output verbatim. Run julius savings for your own report; verbose commands typically compress 60 to 90%.

Three ways it saves.

Commands the agent runs are rewritten through Julius transparently. Your permission rules still apply: denied commands are never touched.

  1. Command-output compression

    Dev commands return a compressed, high-signal version of their output, from 60+ built-in filters.

    git status → julius git status

  2. Application-level interception

    Not just your terminal. A localhost proxy sits inside your app’s own calls to Anthropic and OpenAI.

    ANTHROPIC_BASE_URL=http://127.0.0.1:4141/anthropic

  3. Savings analytics

    Reports tokens saved per command and session. Command numbers are labeled estimates; proxy numbers are exact and provider-reported.

    julius savings

Built for app-level savings.

Every call through the proxy is forwarded verbatim, streaming included, and its provider-reported usage lands in julius savings per app and model.

API usage metering
$ julius proxy serve                 # localhost:4141

$ export ANTHROPIC_BASE_URL=http://127.0.0.1:4141/anthropic
$ export OPENAI_BASE_URL=http://127.0.0.1:4141/openai/v1
$ python my_pipeline.py              # unchanged

  • Runs at the app level

    A localhost proxy sits inside your app’s own calls to Anthropic and OpenAI, catching what CLI-only tools miss.

  • Drop-in, zero code changes

    Point ANTHROPIC_BASE_URL at Julius. No SDK changes, no redeploys, works with any existing app.

  • Free and open source

    MIT licensed, no account, no vendor lock-in. Inspect the filters and run it entirely on your machine.

More from hoop Labs.

Each tool solves one problem, runs locally and needs no signup to try.

See every labs tool

Enforce it across the fleet.

hoop is the same wire-level interception, running centrally, with access policy, data masking and full audit logs across every agent, every app and every LLM call your org makes.

Get started freeView on GitHub (opens in a new tab)

Ask an assistant what hoop.dev does: Claude, ChatGPT, Grok, Gemini, Perplexity, Copilot.

install and hook in
$ brew install hoophq/tap/julius
$ julius init -g