WOZCODE

Cut Claude Code costs by up to 50%

Productivity

Developer Tools

Artificial Intelligence

Visit WebsiteSee on Product HuntTwitter ⧉

One day in and I'm super happy. I can definitively say that the biggest difference I can tell as a user is the number of round trips is significantly less and it feels like Claude understands me much better even though I'm interacting with it in the same way! I'll come back in a week and update this review

Token efficiency is something we’ve been thinking hard about - we use Claude Code heavily for building internal automation and the context window costs add up fast on complex multi-file tasks. The two-command install is the right move; anything that adds friction to the setup kills adoption with dev teams. One question: does this work well with long-running agentic sessions, or is the optimization primarily for single-task completions?

Token efficiency is something we've been thinking hard about — we use Claude Code heavily for building internal automation and the context window costs add up fast on complex multi-file tasks. The two-command install is the right move; anything that adds friction to the setup kills adoption with dev teams. One question: does this work well with long-running agentic sessions, or is the optimization primarily for single-task completions?

Does WOZCODE change how context is selected, like summarizing files before sending them to Claude?

I've always thought that there was a way to more efficiently save credits for AI agents. Will this also help Claude Code keep it's efficiency and learning for preferred results?

First I really dig the design of the site! Second cutting the cost of CC, sounds like a dream haha, we can ship more! Thanks for a great product!

Must use product. Been using it for the past little while, and I can finally calm my token-nerves.

but I wonder, why not just use Pi instead of Claude Code in general? Pi coding agent uses barely no tokens, so token aware users should switch their agent harness, no?

Love the simplicity of it! Wondering how you are measuring how much the session would have cost without WOZ. Is it an estimate or you can somehow figure it out precisely?

We recently refactored our entire platform using Wozcode and it saved us weeks. Setup and auth have been a little messy at times, but overall it’s been a game changer for us at Passive. One of the biggest differences is that we hit Claude limits way less often now, which lets us stay in flow and keep building for much longer without constantly resetting context or starting over. The whole development process just feels cleaner, faster, and more structured. Love the product and look forward to seeing it get even better moving forward!

half of every claude code session is just rereads after edits, ive watched it happen in real time and felt the bill for it. fixing it at the tool layer instead of bolting on another wrapper is the right move.

cool stuff guys !!

hey @ben_collins4
been using this on claude max 20x plan, the token optimisation is real. I hit the $100 free tier limit in a few hours though, which is kind of ironic for a tool that saves tokens. couldn't see what the paid tier limits would be before upgrading, so I paused there. would love clearer upgrade transparency and smoother onboarding.
great job, rooting for you!

The extra 300+ hours per month is wild. How do you think this shifts the balance between solo devs and larger teams? Great idea and Congratulations!

The cost angle is super relevant. Especially once you move beyond experimentation.

I’m currently using Claude Code to build an internal dashboard, and token usage adds up faster than expected.

How are you optimizing for cost without compromising output quality or context depth?

the cost angle is one of the few honest framings out there right now. we run a heavy claude code workload internally and the bill jumps the moment you go from prototyping to actual production loops, so this hit close to home.

two things i would love your take on:

what is the model-routing logic under the hood, is it static rules per task type or something more dynamic that adapts to prompt characteristics? curious because we have found heuristic-based routing starts breaking once heavy tool-calling enters the picture.

and on the 50% number, is that measured like-for-like on the same task graph or an aggregate across mixed workloads? not pushing back, just trying to understand what setup we would need to reproduce something similar internally.

Hey PH! Brad here, Cofounder and CTO @ Woz.

Couldn’t be more excited for you to see what we’ve been building. People always ask: faster, cheaper, better… sounds too good to be true. So how does it actually work?

At the core we’ve improved the way Claude Code optimizes context and all the benefits are downstream from that.

Instead of relying solely on the default tools inside Claude Code, WOZCODE uses a set of custom tools designed for efficiency and context optimization. For example, in vanilla Claude Code, a simple "find and edit 3 files" takes 12+ calls (3× Glob/Grep + 3× Read + 3× Edit + 3x Verify Read). By the final step, it’s reprocessing all prior output as input tokens.

WOZCODE collapses that to 2 calls (1× Search + 1× batch Edit). Context stays small, so every step that follows is cheaper and faster. Over a 30-prompt session, that compounding effect adds up fast.

Happy to answer any questions you may have. You can also check out https://www.wozcode.com/how-it-works to see benchmarks and more technical detail.