PromptForce is in public beta: same AI coding, for less.

If you use Claude Code or Codex all day, you know the two problems. The bill, or the plan limit, arrives faster than you'd like. And every session the agent starts again as a stranger: it re-reads the same files, relearns the same rules and asks you for the same brief.

PromptForce is one desktop app that works on both. It installs two tools that sit beside the agents you already use. You keep the same sign-in, the same code and the same way of working.

Prefex: more work from the AI budget you already pay for

Prefex runs on your machine, between your coding tool and the model. It does a few unglamorous things on every turn:

• Keeps your prompt cache warm and stable, so each turn re-reads what the provider already holds instead of paying to write it again. This is where most of the saving comes from.
• Matches the job to the model, and trims long logs, JSON and repeated file reads before they are sent. Anything it cuts can be fetched back.
• Hands side jobs to free and local models. Summaries, searches and reviews can go to a free pool of 32 models across 10 providers, or to a model on your own machine, within the privacy level you set for each project.
• Shows you the receipts: what each turn cost and what Prefex saved, project by project, with a coaching tip in the chat when it spots something worth changing.

Here is what it measured on our own machine over the last six weeks, across 118,000 requests under heavy daily use of Claude Code:

One machine, one team's workload. Yours will differ, and the dashboard will show you exactly what it is.

ReadyBase: the context each turn needs

An expensive model shouldn't have to rediscover your project every turn. ReadyBase hands it the smallest brief that still gets the turn right: the files, dependents and rules the change touches, with no graph database to run and no stuffing the whole repo into context.

It also scores each repository for how ready it is for agents, and keeps an agent rollout scorecard: which repos took agent commits with no guard on risky files, what each landed commit cost, and how much of the code it wrote is still there. That tells you where to give agents more room and where to keep a human close.

One app to set it up

Download PromptForce for Mac or Windows. Setup finds Claude Code and Codex, turns on Prefex and ReadyBase, and offers the add-ons this computer can run: helper agents, sharper routing, and a local model on Macs with 16 GB or more. It says plainly when something can't install and how to add it later. From then on it lives in your menu bar and keeps both tools up to date from a signed release feed.

Free until it has saved you $500

What leaves your machine

Your prompts and code go only to the models you use. Side jobs stay within the privacy level you choose per project: your own machines, providers that say they don't train on your data, or any provider. Two things are sent to us automatically: an anonymous usage report, which you can turn off, and a daily licence check for the free allowance. No prompt, code or file path is ever sent to PromptForce. The details are on our trust page.

Where this is going

Coding agents arrive brilliant and new, and they leave nothing behind. We think every turn should leave a record you own: what the agent was asked, what it knew, which model did the work, what it cost and whether the change lasted. That record is how agents earn more autonomy on your code, how more of the work moves to models you run yourself, and, one day, the data you tune your own models on. Prefex and ReadyBase are the first two pieces.

Tell us what breaks

This is a beta. Some things will be rough, and we would rather hear about them than guess. Come and tell us on Discord or open an issue on GitHub.