Skip to content
Comparisons

Don’t choose free-pi for private or client code: every session trains models and may be sold

free-pi’s own consent text says your prompts, code and files train models and may be sold to third parties, with no opt-out once you accept.

In short

  • Choose Freebuff, not free-pi, for private or client code. free-pi trains on every session and says your code may be sold.
  • Every session is training data. free-pi’s ads and privacy doc says your prompts, code, files and responses "are saved as traces and used to train models" and "may be shared or sold to third parties".
  • No opt-out, no delete button. After you accept there is "no opt-out, except where the law gives you one", and "no self-service deletion yet".
  • An ad partner sees your recent conversation. Some ads come from a partner that "receives excerpts of your recent conversation", and the privacy policy keeps your full IP address "indefinitely".

free-pi (freepi.ai) is a free, ad-supported terminal agent built on Mario Zechner’s Pi harness, run by DappHero Corp. It is upfront about its data terms, and those terms are the reason not to open private or client code in it. Here is what free-pi’s own pages say, checked on November 17, 2026.

Every session becomes training data

free-pi’s ads and privacy doc summarizes the consent screen you accept on first run. Your coding sessions, meaning prompts, code, files and model responses, "are saved as traces and used to train models". They "may be shared or sold to third parties".

The privacy policy says the same, and notes that traces can include "proprietary code", credentials, API keys and passwords if they show up in your session. free-pi tries to strip personal information first, but says it "cannot promise" a trace is perfectly clean.

No opt-out and no delete button

After you accept, there is "no opt-out, except where the law gives you one", and there is "no self-service deletion yet"; you email the company to ask about your data. If you decline the consent screen, the CLI exits, so using free-pi at all means agreeing to this. free-pi’s own docs tell you not to put secrets or code you can’t share into a session.

An ad partner sees your recent conversation

  • Some ads come from an advertising partner that "receives excerpts of your recent conversation" to pick an ad (ads and privacy).
  • free-pi keeps the full IP address you connect from "indefinitely" (privacy policy).
  • Redacted excerpts of your prompts go to a third-party AI service that labels how you use the product.

An early experiment holding your code

free-pi’s homepage calls it "an early experiment" and tells you to "expect it to evolve". Its privacy docs call it "in alpha and testing, with no warranty or uptime guarantee", and its models and limits doc says the daily allowance "is set by the server and can change". That is a lot of uncertainty to attach to code that belongs to a client or employer.

When free-pi is still the right call

If your code is public or a throwaway, training on it costs you nothing. free-pi also runs once with no global install (npx free-pi-cli@latest), suits people who already like the Pi harness, and offers abliterated Defense models for authorized security research.

Choose Freebuff instead

Freebuff is free and ad-supported too, so it uses data as well. The difference is that you choose. Only models labeled "May use data for AI training" are used for training, and Freebuff may share or sell that data, so for private or client code you pick a model without the label. Freebuff also uses your messages to personalize ads and train the models that pick ads. The details are in our privacy policy.

  • A published allowance. Up to 30 hours a day in the US, and 21 in countries such as Canada, the UK and Germany, in hour-long sessions with unlimited messages and no monthly cap. No card and no API key.
  • Models that don’t touch it. The free picker recommends DeepSeek V4.1 Flash, and right now Solar Mini 4 and Solar Pro 4 don’t use the allowance at all.
  • It runs where you work. In the terminal with npm install -g freebuff, as desktop apps for Mac, Windows and Linux, in the browser with the web workspace, and on iOS and Android.
  • 9 specialized subagents, including file-picker, code-reviewer, browser-use and thinker. The agents are open source on GitHub, and Freebuff reads your AGENTS.md and CLAUDE.md.

Freebuff has been live since February 2026, is backed by Y Combinator, and more than 850,000 developers have signed up. Its CLI was downloaded 106,189 times from npm in the 30 days to November 16 (npm API), against 20,720 for free-pi-cli. For the full side-by-side, read Freebuff vs free-pi, and for the privacy terms in more detail, Is free-pi safe for your code?.

Try Freebuff for free

No card, no API key. `npm install -g freebuff`

Get Freebuff

Questions

Should I choose free-pi or Freebuff for private or client code?
Freebuff. free-pi’s consent screen says every session is used to train models and may be shared or sold, with no opt-out once you accept. On Freebuff only models labeled "May use data for AI training" are used for training, so you can pick one without the label. Freebuff also uses your messages to personalize ads and train the models that pick ads.
Is free-pi free?
Yes. free-pi is free with a GitHub sign-in and no card, paid for by terminal ads and by training on your sessions. It also sells optional $5 and $10 credit packs. It doesn’t publish the size of its daily allowance.
Does free-pi train on my code?
Yes. Its consent screen says prompts, code, files and responses are saved as traces, used to train models, and may be shared or sold to third parties, with no opt-out after you accept except where the law gives one.
Is free-pi worth it?
For public or throwaway code, it can be: it is free and runs with one npx command. For code you aren’t allowed to share, its own docs tell you not to put it in a session.
What is the best free alternative to free-pi?
Freebuff. It is free with no card or API key, publishes its allowance (up to 30 hours a day in the US, in hour-long sessions with unlimited messages), and lets you pick models that aren’t labeled for AI training.