Same coding agents.
Your limit lasts longer.

Probe0 runs under Claude Code, Codex and Cursor on your machine. It cuts what they burn through, so your plan goes further, and shows you what every call used.

Download for MacHow it worksBenchmarks

Or install it from your terminal:

$ curl -fsSL https://probe0.xyz/install | bash
All traffic · this machine
Today / cap$18.44 / $25.00
  • claude-code · 41822sonnet-4.6 · $0.0732Upstream
  • codex · 41903qwen3:8b · local · $0.0000Local
  • claude-code · 42011opus-4.7 · blockedCapped

1proxy slot

Your machine has one, and every tool wants it. So you pick an inspector, a cache or a router, and go without the rest.

93%of the 1-hour cache

Claude Code pays for an hour of cache it then uses for five minutes. Probe0 buys the five minutes.

92%of your usage is cache

51% reads, 41% writes, across 85,569 real turns. Most of what you burn goes on the cache, not on your prompt.

§ 01

Four ways your limit lasts longer

It runs under the tools you already use, and any step can be turned off. On 85,569 real turns it cut what they used by 12.8% with no model swaps, and 24% with swaps on.

  1. 01

    Intercept

    Install it once. Every agent you run goes through it — Claude Code, Codex, Cursor.

  2. 02

    Route

    Easy work goes to a model on your own machine, so it never counts against your plan. Hard work still goes to the best one. A weak local answer gets sent up automatically.

  3. 03

    Reuse

    Ask the same thing twice and the second answer never touches your limit. Two agents asking at once only use it once.

  4. 04

    Account

    Every call gets written down: model, tokens, time, which agent, and how much of your plan it used. It warns you at 80%.

02 · route

Easy work doesn't need the best model

The same picture from a cheaper model, then from one on your own machine.

opus-4.784 colsupstream

03 · reuse

Two agents ask at once. Only one counts.

The second request joins the first one's stream. Same answer, counted once.

claude-code · 41822codex · 42113$0.0732upstream$0.0000joined 41822

§ 02

Turn on only the parts you want

Ten switches. Each reports what it saved, so nothing stays on out of faith.

Send less

  • Strip Compression

    Shrinks tool output before it counts against you.

  • Caveman + Ponytail

    Shorter replies and leaner code. Each tunes Light to Heavy.

Send it somewhere cheaper

  • Local Routing

    Whole classes of work run on your own model.

  • Cloud Routing

    Easy turns go to a cheaper model in the same family.

  • Effort step-down

    Same model, less thinking, on easy turns.

Don't send it at all

  • Exact Cache

    Repeat questions never hit the network again.

  • Semantic Cache

    Near-identical questions too, held to a strict match floor.

  • Duplicate Collapse

    Two agents ask at once, it counts once.

Know what it used

  • Spend Guard

    Warns at 80% of your plan. On a pay-per-token key, it pauses.

  • Recording

    Every call, the agent behind it, usage you can check.

§ 03

Pay for less output, keep the cache

Rewrite a prompt and you lose the copy your provider is already holding. Probe0 only shrinks what your tools send back, so that copy keeps paying off.

Strip Compressionon by default
Bash output comes back about 30% smaller, the turn it arrives. Old file reads shrink later, once the provider has stopped holding them. Worth about three points of your usage.
Caveman + Ponytailon at Light
Caveman shortens replies to you. Ponytail has the agent write the least code that does the job, keeping validation and error handling. Each has its own strength. Posts and emails your agent writes stay normal. Savings are shown as estimates.
$ pnpm vitest run tests/unitwhat the model is sent14 lines · 1,362 chars

A 96-line test run folded to 14 lines by Probe0's shell rules. Passing tests are collapsed behind a marker; the summary is kept.

raw 96 lines · 10,247 charsone real run, from the Probe0 repo

§ 04

One slot, and anyone can build in it

Long term, this is meant to be the layer other people build on.

Only one thing can hold the slot
Your machine has one proxy position. An inspector, or a cache, or a router. Not all three.
So the slot should be programmable
Probe0 takes the position and then opens it up. A module sees a request and decides what happens to it.
A module asks for what it needs
Read the details, change the body, reach the network — and nothing else. You approve that list when you install it.
Everything published is reviewed
Asking up front is what makes review possible: only what a module asked for has to be checked.

§ 05

The call that never touches your limit

Plenty of agent work doesn't need the best model on the market.

Your models, your hardware
Point Probe0 at a model already running in Ollama or LM Studio. It becomes another tier.
Or a cheaper cloud model
Step down to a cheaper model in the same family instead. Requests carrying tool calls take the safe step only.
You set the share, not a rule
Both are a setting: how much of your traffic takes the cheaper path. Left at the default, about 11% of turns.
Your code never leaves your machine
The proxy, the cache and the record of every call all run here.
Both asked for opus-4.7routed on this machine

  1. codex · 41903 "rename a variable"no tool calls

    probe0qwen3:8b @ localhost:11434$0.0000

  2. claude-code · 41822 "fix the failing test"carries tool calls

    probe0sonnet-4.6 @ api.anthropic.com$0.0732

§ 06

See how much of your plan is left

Probe0 measures what every call uses against what your plan includes. It warns you at 80%, so you can move the easy work to a cheaper path before you run out.

It works the other way too. Barely touching your plan? It tells you which tier would do.

Last 30 days · this machine
Plan used
86%
Answered from cache
2,317
Duplicate calls collapsed
1,842
Cache writes avoided
3,106

On pace for 86% of your Max 5x plan. Warned you at 80%.

Try it on your own agents

Free beta on macOS. Sign in, download, and it starts counting on the next run. Nothing gets uploaded.

Download for MacmacOS · Apple silicon and Intel.

Or from your terminal. One command installs Probe0, turns it on, and shows what your last 30 days of Claude Code cost. Undo it with probe0 uninstall.

$ curl -fsSL https://probe0.xyz/install | bash

Ask an AI instead

Opens a chat that already knows Probe0 and its benchmarks.