Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

Analysis · September 17, 2026 · repo

PRAXISTSapient IntelligenceFair Sourcemulti-agentagent-harnessregulationai-skills

PRAXIST Hit #1 on GitHub Trending Under a License Called Fair Source That Never Converts to Open Source

The autonomous R&D agent from Sapient Intelligence is source-available, revenue-gated, and stamps its name on anything you publish. Read the file, not the badge.

The top repo on Trendshift's daily board this morning is an autonomous research system with 5.7k stars, a paper on arXiv, and a license file that opens with the words "Fair Source License Agreement (Version 1.0)." Fair Source is a real thing. It has a website, a definition, a governance README, and a list of approved licenses. Sapient's license is not on that list, and it fails the definition's third test on its face.

That gap is the story. Not because PRAXIST is hiding anything (the README FAQ describes the terms accurately, and the license file is 300 lines of plain English), but because a license name is doing work that the license text does not do, and GitHub's star counter cannot tell the difference.

What PRAXIST is

sapientinc/PRAXIST calls itself "an autonomous research system for measurable, computer-executable research." You point it at a project that already runs and has a metric, and it launches parallel "research peers" that propose and test changes across generations, with evaluators turning results into evidence and a planning panel synthesizing that evidence into the next generation's agenda. The install is one line:

python3 -m pip install --index-url https://pypi.org/simple "praxist[agents,codex]" && praxist setup --interactive --install-skills codex

After that you open Codex in the project root and invoke a $praxist-takeover skill, which inspects readiness, builds a task harness with metric directions and baseline provenance, and launches a detached run. There is a Claude Code path too. A Codex subscription works without an API key. The README's FAQ is unusually candid: it says PRAXIST "does not guarantee a specific metric improvement" and will hand you a negative-result evidence package if a run fails. The latest release is 0.5.0, dated August 27, and the paper is arXiv 2608.25955.

This is the kind of tool that gets starred in bulk by people who will never run it, because "personal R&D team" is a strong pitch. That is fine. The question is what the star is endorsing.

What Fair Source means, and what this file says

fair.io publishes a three-part Fair Source Definition. The software is publicly available to read; it allows use, modification, and redistribution "with minimal restrictions to protect the producer's business model"; and it "undergoes delayed Open Source publication (DOSP)." The site calls the first and third points "bright lines." DOSP is the part that separates Fair Source from plain source-available: after a planned period, the code goes out under an OSI-approved license, so if the company dies or turns, the community can carry the code forward. The three licenses fair.io recommends (FSL, FCL, BUSL) all convert.

Now read Sapient's LICENSE.md. It is dated August 28, 2026 at the bottom. There is no conversion clause. Nothing in the file becomes open source at any date. What it does contain:

A revenue threshold. The free license is valid only while your aggregate annual gross revenue, including affiliates, is under $1,000,000 USD. Cross it and the free license "shall automatically lapse," you have 30 days to notify Sapient, 60 more to negotiate a commercial license, and if no deal lands by day 90 you must "cease using the Software and delete all copies." Failing to send the notice is defined as an "incurable material breach."

A reporting obligation. Every twelve months Sapient may require a signed revenue statement, due within 30 days, and may terminate if you refuse.

An output attribution clause. Internal use needs no credit. But if you "publicly release, distribute or otherwise make available to any third party any output generated by the product," you must keep "Praxist by Sapient Intelligence" in "a reasonably prominent position," and the obligation "shall continue to apply even if the Licensee modifies, edits or otherwise reprocesses the output." For a tool whose output is a research result, that is a citation requirement enforced by contract rather than by academic norm.

A distribution ban. You cannot redistribute the software or a modified version as a standalone product without written consent. You can offer its functionality inside your own service as long as you do not sell it under the Praxist name.

Singapore law, SIAC arbitration, and a $100 liability cap if you used it for free.

None of this is unusual for a commercial source-available license. It is roughly the shape of a Business Source License with the conversion date removed and an attribution rider added. What is unusual is the name on the tin.

The text that gives away where the license came from

Two definitions in the file do not describe PRAXIST. Section 1.1.5 defines "Generated Output" as "any model parameter updates, model weight files, inference outputs or optimization results." Section 1.7, the warranty disclaimer, states "The Software is an AI model tuning tool, and its output may contain bias, errors or inaccuracies. The Licensee should independently verify results prior to deploying any model tuned through the Software."

PRAXIST is not a model tuning tool. Its README says it "contains no task-specific scientific assumptions" and that the task project, not PRAXIST, owns "executable code, evaluator, metrics, baselines." The license text reads like it was drafted for a different Sapient product and reused, which is a common thing to happen and a useful thing to notice, because it tells you the attribution clause was written with model weights in mind and is now being applied to whatever PRAXIST produces: a patched kernel, a tuned hyperparameter, a paragraph in a paper.

I do not think this is malicious. I think it is a company that wanted the word "fair" and did not read fair.io's definition, or read it and decided the name was worth more than the conversion clause. Either way, the effect is the same: a reader who knows what Fair Source means will assume a future they are not getting.

Why this matters more for agents than for databases

Source-available licenses were mostly fought over infrastructure: databases, queues, observability. The buyer was a platform team, and the license question was "can a cloud provider resell this." Autonomous research agents change the buyer and the question. The buyer is a researcher or a small lab, the output is a result rather than a service, and the license attaches to the result.

The revenue gate in PRAXIST is measured on gross revenue of the licensee and affiliates. A three-person startup under a $1M line is fine. A university spin-off using it through its tech-transfer arm is explicitly excluded from the academic exemption. A team inside a larger company is over the line on day one, because the affiliate definition rolls up to anything that controls more than 50% of you. That is the case nobody stars for: the person at a mid-size company who runs pip install praxist on a Friday afternoon has, per section 1.3.3, started a 30-day clock.

And GitHub shows all of them the same number.

Put this into practice

Read the LICENSE file, not the badge, and not the README's summary. The README FAQ here is accurate, but a FAQ is a summary "and if it conflicts with the license agreement, the terms of the license agreement control," as this one says of itself. Fetch the raw file. If it has a name you recognize, check that it is actually that license: fair.io lists its licenses, OSI lists its licenses, and a name-alike is not a match.

If a license has a revenue threshold, find the definition of revenue. Gross versus net, licensee versus licensee-plus-affiliates, and the annualization rule for young companies all change who is over the line. PRAXIST uses gross, includes affiliates, and annualizes.

If a license has an output clause, decide before the first run what you plan to publish. PRAXIST's attribution requirement is cheap to satisfy (a footer or metadata line) and expensive to discover after a paper is out.

For PRAXIST specifically, if you are under the threshold or academic and you want to try it: the Codex-native path needs no API key, the takeover skill refuses to run on a project that does not already execute, and the README recommends starting with a small representative workload before scaling generations. Run praxist doctor first. Keep the run artifacts separate from your repo, which it does by default.

And if you are shopping for this category, note the alternatives with real licenses. The Trendshift board the same morning had Paper2Agent under MIT and Memoh under AGPLv3, each with their own catches, but each with a file that means what its name says.

Honest limitations

I have not run PRAXIST. Everything above comes from the README, the license file, the release feed, and the arXiv listing, all fetched with cache-busting on September 17. I cannot tell you whether the research loop works, and the project's own FAQ is careful not to promise that it does.

Sapient may well intend to add a conversion clause, or may consider "Fair Source" a generic phrase. The company's own FAQ uses the more precise term "source-available" in its second sentence, which suggests someone there knows the difference. I have not contacted them.

The 5.7k star figure is a cache-busted shields.io read from this morning and will be stale by the time you read this. Trendshift's #1 is a momentum rank, not a size rank.

Nothing here is legal advice. The license is governed by Singapore law and I am a writer with a text editor.

Closing

A star is a vote for a name. PRAXIST earned its votes with a good pitch and a real paper, and it will keep earning them, because the license file is the last thing anyone opens. The fix is not a campaign against source-available software; companies can license however they like. The fix is to open the file, find the clause that converts, and, when it is not there, stop calling it Fair Source.

You have the URL. It takes two minutes.

Sources: sapientinc/PRAXIST README and LICENSE.md (raw, cache-busted, September 17, 2026); fair.io About and Licenses; arXiv 2608.25955; Trendshift daily board read 7:12am ET.

Medium metadata

  • Title: PRAXIST Hit #1 on GitHub Trending Under a License Called Fair Source That Never Converts to Open Source
  • Subtitle: The autonomous R&D agent from Sapient Intelligence is source-available, revenue-gated, and stamps its name on anything you publish. Read the file, not the badge.
  • Tags: Open Source, AI Agents, Software Licensing, AI Research, Startups
  • Canonical: fervorai.dev (import from the published URL)