Offloadby Boxwright

Use less credits with your home box AI.

Offload is a small daemon for people who own a GPU box and pay for Claude. You give it one line about a change you want. The model on your own GPU does the typing. Claude only plans, reviews, and steps in when the local model stalls. You get a pushed branch, a report, and a message when it needs a decision.

Check your box first Read the code
real session, recorded

    

  


  

    
One real job: write a 33-test suite for Offload itself, 446 lines, 5 filesdrawn to scale

    
    
The brass sliver is Claude: 36 seconds, 2 calls, one to plan and one to review. The job's report

  


  

    
Before you install anything

    
Will it run on your box?

    
Most machines cannot run this, and it is better to know now. This uses the same rules as the preflight script, which reads your actual hardware and installs nothing.

    

      

        

          

        

          

        

          

        

          

      

      

        

        
    
        
    
      
    
    
    
  
    

  
    
    
    One harness, two brains
    
    
    How it works
    
    
    Both workers are Claude Code in print mode, inside the same locked-down container. The local one talks to a small proxy in front of your model server, so there is no second agent to install. Brass marks the only places the paid model is used.
    
    
    
      
        
          
          
        
        
        one lineoffload add "…"
        
        
        intake + planClaude · once each
        
        
        steps: your GPU, in the sandbox
        a fresh session for every step
        tests after each step, in a container
        with no network; failures are fed back
        
        red tests → try again
        
        
        one rescueClaude · only when it stalls
        
        
        reviewClaude · once
        
        
        one commitpush a branch
        report · messageby the engine
      
    
    
    
    
      
    Rate limits are its problem
    On a limit it reads the reset time, waits, and resumes the same session. In our test the resumed review cost a tenth of a fresh one.
    
      
    A budget you set
    It paces Claude use evenly across your week and keeps a share for your own sessions. Ahead of pace, the paid steps wait and the local ones continue.
    
      
    Three questions, ever
    It asks before spending money, publishing, or deleting. Everything else it decides, and writes down in the job's report.
    
      
    Reports, not vibes
    Every job leaves a page: what changed, what it cost, who decided what, and whether the paid model had to rescue a step.
    
    
    
  
    

  
    
    
    Measured on real jobs, evidence in the repository
    
    
    Receipts
    
    
    
      
    
        
    JobYour GPUClaudeClaude's share of the time
    
        
          
    Fix a bug in a small repo, the built-in offload demo58 s · 3 sessions20 s · 2 calls
    
          
    Write a 33-test suite, 446 lines, 5 files740 s · 7 sessions36 s · 2 calls
    
          
    Move 13 hardcoded settings into a config layer, with tests868 s · 5 sessions68 s · 2 calls
    
          
    Turn a 600-line script into a package with a CLI4,294 s · 6 sessions104 s · 2 calls
    
            
      
    
    
    
    
    The last three jobs were Offload working on its own source code. On a five-task calibration set the local model and Claude Sonnet both scored 5 of 5; the same small tasks done entirely by Claude re-read 130,000 to 190,000 cached tokens each. Hardware: one RTX 5090, Qwen3.8-27B. Your numbers will differ; the runners are in the repo so you can measure your own. Calibration results · Hardware page
    
  
    

  
    
    
    
      
    
        
    Who should not install this
    
        
    What it cannot do
    
        
      
          
    • Make a small GPU run a big model. Under 22 GB of VRAM there is nothing to install yet.
      • 
          
    • Use a 32 GB laptop as the host. We measured one: fine at writing, a tenth of the speed at reading, so every session would wait about a minute before its first word.
      • 
          
    • Get vague or sprawling tasks right on its own. It escalates when the local model stalls, and the report says when it did.
      • 
          
    • Run on Windows, or with a Mac as the host. A laptop can drive a Linux box over SSH.
      • 
        
    
      
    
      
    
        
    Models editing code while you sleep
    
        
    Safety, in short
    
        
      
          
    • Every model runs in a container, unprivileged, behind an outbound firewall it cannot change. If the firewall does not start, the worker does not run.
      • 
          
    • The local model's container never holds your Claude login, and cannot reach the internet at all. The paid one can reach Anthropic and nothing else.
      • 
          
    • Your project's tests run with no network at all, because tests are code the model just wrote.
      • 
          
    • An allowlist of tools, not a blocklist. No push, sudo, ssh, docker, curl or pip.
      • 
          
    • The engine makes the one commit and the one push, to a branch, never to your default branch by itself.
      • 
        
    
        
    The full safety model, including what it does not protect against.
    
      
    
    
    
  
    

  
    
    
    Two commands, and the first installs nothing
    
    
    Install
    
    
    
      
    
        
    1 · Ask your machine
    
        
    Read-only. It reports your GPU and memory, Docker, the NVIDIA container runtime, disk space, and the Claude CLI, and says plainly what is missing.
    
        
    $ curl -fsSL https://raw.githubusercontent.com/boxwright/offload/main/install/preflight.sh | bash
    
      
    
      
    
        
    2 · Install
    
        
    It tells you the size before every download (the model is 17 to 20 GB), offers speculative decoding with a recommendation based on your GPU memory, and never sees your Claude login: you create your own token with claude setup-token.
    
        
    $ curl -fsSL https://raw.githubusercontent.com/boxwright/offload/main/install/install.sh | bash
$ offload doctor     # checks Docker, the sandbox, the model, the token
$ offload demo       # a two-minute job on a sample repo, so you can watch it work
    
      
    
    
    
  
    

  
    
    
    Help wanted
    
    
    Your hardware
    
    
    Everything measured here comes from one machine: an RTX 5090 with 32 GB, two models. Every other line on the hardware page is an estimate, and says so. If you own a Linux box with an NVIDIA card of 24 GB or more, the read-only check above takes a minute, and its output in a hardware report turns an estimate into a measurement. Nobody has measured any 24 GB card, or the 20 GB download on a machine that did not have the file. A pull request that adds a row to docs/hardware.md is the best kind: card, driver, model and quant, context, VRAM after load, tokens per second, date, and what did not work.
    
  
    

  
    
    
    Built in the open, mistakes included
    
    
    The story so far
    
    
    Four things that went wrong while building it, each with a number behind it.
    
    
    
      
    One word in my prompt cost me 415,000 tokens.The daemon looked for the word STUCK. The local agent echoes its instructions, which said "say STUCK".
    
      
    My laptop fell asleep. The job didn't.A 73-minute refactor finished, was reviewed and pushed. I found out two days later.
    
      
    My AI sandbox had a door I forgot to lock.Every model was sandboxed. The tests, which are code the model wrote, ran on the host.
    
      
    It kept calling for backup. I had told it to stare at the code first.One sentence to the planner: 38 cents and a rescue became 13 cents and none.