Atropos

license

MIT(OSI)

source

public

scope

async RL-environments framework + trajectory API

trainer-integrations

Axolotl/Tinker

core-gated

ungated

MIT-licensed, an OSI license, with the source public and no managed or closed core. It sits in this category as the RL-training-environment layer beneath RLHF, RLAIF and GRPO training runs.

• https://github.com/NousResearch/atropos recorded 2026-08-08

  Repo page carries the label "Public archive" and the banner "This repository was archived by the owner on Jul 4, 2026. It is now read-only." Embedded repo JSON reads "isPrivate":false and "isArchived":true alongside "license":{"spdxId":"MIT","name":"MIT License"}. The rendered README says "Atropos is an environment microservice framework for async RL with LLMs." and its Navigating the Repo table lists the product's own parts in-repo: atroposlib/ ("Core library containing base classes and utilities"), environments/ ("Collection of ready-to-use RL environments"), example_trainer/ ("Example training scripts and configurations"); installation is "pip install atroposlib" or an editable install of the same tree. So the published repo is the runnable product: source public. For core-gated I extracted the rendered README text out of the saved body (21,019 characters) and searched it for enterprise, pricing, paid, commercial, subscription, premium, cloud tier, license key, sign up, contact sales and managed: zero hits on all of them except one occurrence of "managed", which is the ManagedServer class documentation link. Nothing in the repo is withheld for a paid tier.

• https://raw.githubusercontent.com/NousResearch/atropos/main/LICENSE recorded 2026-08-08

  The LICENSE body in full, 1,070 bytes. It opens "MIT License", carries "Copyright (c) 2026 Nous Research", and grants permission to deal in the software "without restriction" including use, copy, modify, merge, publish, distribute, sublicense and sell, subject only to the notice-retention condition and the standard warranty disclaimer. No field-of-use, non-commercial, competition or delayed-conversion clause anywhere in the text, so the recorded license:MIT(OSI) holds and resolves to the osi tier.

• https://api.github.com/repos/NousResearch/atropos recorded 2026-08-08

  Repository JSON: "private": false, "visibility": "public", "archived": true, "disabled": false, "spdx_id": "MIT", "pushed_at": "2026-07-04T17:39:01Z", "open_issues_count": 0, default_branch main, created 2025-04-29. Corroborates the landing page on both the public-source and the MIT reading, and independently confirms the archived flag the warehouse GitHub signal reported.

• https://portal.nousresearch.com/ recorded 2026-08-08

  The vendor's paid surface, and the pricing page the software ladder says core_gated needs. It sells a Nous Portal subscription built around the Hermes agent: nav sections Overview, Hermes Cloud, Models, Plans, Resources, a Pricing heading, and copy about hosted tools that "come bundled with your subscription" and "bills to the same credits as your models". The string atropos (case-insensitive) does not occur anywhere in the saved body of 647,329 bytes, searched over the raw HTML rather than the rendered text so that script-embedded payloads were included. Nothing from the Atropos repo is sold as, or withheld for, a Portal tier: core-gated ungated.

• https://nousresearch.com/ recorded 2026-08-08

  Vendor home page. The product nav is Hermes Agent, Nous Portal, Psyche, Hermes 4, Chat, Simulators, Releases, Careers, Shop, Blog. Searching the raw saved body of 76,131 bytes case-insensitively for atropos, pricing, enterprise, subscri and plan returns zero hits for every one of them, so Atropos is not a commercial offering of this vendor at all and there is no tier for a feature to be withheld into.

atroposlib records 157 PyPI downloads in the trailing 30 days, well inside the under-10K band -- early, niche research adoption. The repository was archived read-only on 2026-07-04, so the figure is unlikely to grow.

• https://pypistats.org/api/packages/atroposlib/recent recorded 2026-08-08

  The whole 106-byte body is {"data":{"last_day":10,"last_month":157,"last_week":40},"package":"atroposlib","type":"recent_downloads"}. 157 downloads in the trailing 30 days for the package the README installs, which bands at level 1 (<10K) with a wide margin to the 10K level-2 floor. Fetched 200 on the first attempt, no retry needed.

• https://pypi.org/pypi/atroposlib/json recorded 2026-08-08

  info.version 0.4.0, summary "Atropos: An Environment and Rollout handler for LLM RL", requires_python >=3.10, five releases (0.1.0, 0.2.0, 0.2.1, 0.3.0, 0.4.0) with the 0.4.0 files uploaded 2026-03-10T04:21:18Z. Two things matter here. First, the record's downloads object is present and filled with sentinels rather than absent: "downloads":{"last_day":-1,"last_month":-1,"last_week":-1}. It carries no figure; a reader that treats -1 as a count would band this product at nothing, and a reader that treats the key's presence as a figure is also wrong. Second, the 25,750-character long description is the repo README and links github.com/NousResearch/atropos, which is what ties the atroposlib download count to this product, since sources/products/atropos.yaml records no pypi artifact of its own.

• https://api.github.com/repos/NousResearch/atropos recorded 2026-08-08

  stargazers_count 1343, forks_count 396, open_issues_count 0, archived true, pushed_at 2026-07-04T17:39:01Z. Corroboration only, and consistent with a level-1 band: attention at the low-thousands of stars on a repo that has stopped taking changes. It is no longer the signal the score rests on.

Capable, well-architected RL-environments layer; mid-tier as a fine-tuning enabler (it orchestrates RL rather than implementing large-scale training itself). Vendor-reported gains (4.6x parallel-task, 2.5x finance) not independently benchmarked.

• https://github.com/NousResearch/atropos recorded 2026-08-08

  The rendered README carries the feature matrix the band rests on. Its environment table lists "Dataset environments" ("GSM8K, MMLU, Custom HF Datasets"), "Online environments" ("Blackjack, Taxi, Text-based games"), "RLAIF and RLHF" ("LLM Judge/Reward Models"), "Multi-Turn RL", "Code Execution" and "Multimodal". A "Trainer Integrations" section covers Axolotl and Tinker, and the diagram caption notes the trainer and inference engine are not included with the atropos package, with the in-repo example trainer described as a reference example. Tooling includes "view-run", process, and "Offline Data Generation" via atropos-sft-gen and atropos-dpo-gen, over OpenAI-compatible endpoints. The vendor's own results table reports a "Berkeley Function Calling Benchmark" parallel-task improvement of "4.6x" and a simple-task move from 21 percent to 51.75 percent, both self-reported and not independently benchmarked. "GRPO" does not occur anywhere in the README.

• https://api.github.com/repos/NousResearch/atropos/releases recorded 2026-08-08

  Four published releases, newest "tag_name": "v0.4.0" with "published_at": "2026-03-10T04:20:36Z" and "prerelease": false; the others are v0.3.0 (2025-07-16), v0.2.1 and v0.2.0. Together with the repo being archived on 2026-07-04 this establishes that the feature set is frozen at 0.4.0, which is why the capability band is a description of a fixed matrix rather than a moving target.