EVA-OS is a platform announcement with almost no code in it

The repository holds documentation, not a library

The top level of AutoArk/EVA-OS contains four entries: a .opencode directory, LICENSE, README.md and docs. There is no source tree, no module manifest, no test suite and no build file. That single fact reframes everything else the README claims.

The project describes itself as a real-time multimodal AIOS for hardware, and its headline is a platform launch dated 30 July 2026, with a release tag release/v2.0.0 for EVA-Platform V2.0.0 on 2026-07-31. The licence is MIT and the README reproduces the standard grant: you may use, copy, modify, merge, publish, distribute, sublicense and sell copies provided the copyright notice travels with them. An MIT grant over a directory containing only prose is worth noticing, because it tells you what a future code drop would be licensed under, not what you can install today.

The last push was on 2026-08-03. The repository is not archived, but the gap between that date and the platform launch is what a reader should hold onto when judging maturity.

Five capabilities, three of which live outside GitHub

The README breaks the platform into numbered pieces, and separating them by where the code lives makes the structure clear.

EVA SDK is the developer-facing piece: full-duplex audio and video with low-latency communication, barge-in and interruption, emotion recognition, and device command dispatch. The README says the SDK targets apps, IoT and embedded devices, currently in TypeScript and Python, with Flutter, ESP32 and what it calls a domestic chip family described as work in progress. The examples live in a separate repository, AutoArk/eva-sdk-examples.

EVA CLI is the terminal side: creating and managing projects, applying for and managing API keys, checking resources and usage, and configuring application settings. The README frames it explicitly as a toolchain an AI Agent can also execute, within an authorised scope.

Model Gateway is the abstraction layer over ASR, TTS and LLM providers, with the stated aim of switching models through one parameter without touching business code. On-device inference, the fourth capability, is described as part of the commercial edition, running ASR and TTS on the device using local compute.

The gateway diagram is the whole integration story

The only technical diagram the README offers is a three-line text flow, and it is worth quoting because it is the entire architectural claim:

your application -> EVA Model Gateway -> EVA / mainstream model providers
                                             ASR / TTS / LLM

The arrow structure says the gateway sits between your code and whichever provider serves the request. Authentication, call format and usage accounting are unified at that layer, per the README, and the practical benefit it names is that changing a model does not require rewriting the application.

What the diagram does not show is the transport. The SDK bullet on full-duplex audio names low-latency realtime communication, and the repository topics list WebRTC, but neither the README nor the docs directory gives a protocol, a port, an SDK method name or a sample call. Anyone assessing integration cost is working from the feature list, not from an interface.

Getting started means signing up, not running a command

Neither quick-start path in the README contains a shell command. The first path is the AI Native workflow: register on the platform site, obtain the EVA Agent Skill and install it into your AI coding tool, pick a demo or template, describe the application in natural language, and let the agent scaffold the project, wire the SDK and configure it.

The second path is the SDK path: create or use a default project on the platform, apply for a project API key, pick the SDK for your environment, run an example against your app or device, then configure ASR, TTS and LLM providers as needed. Step three is listed as choosing an SDK by environment, but the sentence that would name the packages is missing from the README.

So the honest tutorial is short. Get an account, get an API key, and read the SDK documentation in the examples repository. There is no offline trial path, no self-hosted mode described here, and no statement of what the free monthly quota actually covers.

The leaderboard claim is scoped to the commercial edition

One performance claim appears, and it needs careful reading. The README states that as of the platform launch, the in-house ASR model, in its open-source version, ranked first on the Hugging Face Open ASR Leaderboard among full-size models, and it embeds a screenshot from docs/images as evidence.

Two qualifications sit next to that sentence. First, the ranking is scoped to a leaderboard configuration and a model size class, both of which move over time, so the statement is a snapshot of a specific date rather than a durable property. Second, the capability paragraph above it describes on-device ASR and TTS integration as belonging to EVA 商用版, the commercial edition, while the leaderboard claim refers to the open-source version of the ASR model. Those are two different distributions, and the README does not say which artifact a developer gets when they create a free project.

There is no latency figure, no word error rate and no hardware configuration anywhere in the repository, so no comparison against another ASR engine can be made from what is here.

Where the fit is narrow

The stated use cases are companion hardware such as smart toys and desktop robots, voice devices including speakers and in-car assistants, realtime AI applications such as language practice and meeting assistants, IoT and embodied systems, and products needing local ASR and TTS with low inference latency.

That list points at a specific buyer: a hardware team that already has a board and needs speech in and out plus device commands, and that would rather buy a managed path than assemble a realtime audio stack. It is a poor fit for a web developer who needs a speech API, and a poor fit for anyone requiring the SDK source to be auditable, since this repository does not carry it.

There is also a maintenance signal to weigh. The platform version is 2.0.0, the second release tag in the repository's list is a resource release dated 2025-12-04, and the last push came on 2026-08-03. A repository that is mostly a landing page can change its README without any code review, so treat its cadence as marketing cadence until code appears in the tree.

Editorial conclusion

Treat EVA-OS as a commercial platform evaluation rather than an open source dependency: the TypeScript and Python SDKs it points you to live in a separate examples repository, and the on-device ASR and TTS models it advertises are described as part of the commercial edition. Before building on it, open the eva-sdk-examples repository and confirm that the full-duplex audio path, the emotion recognition output and the API key model work with hardware you actually have, because none of those are demonstrated in this repository.

Frequently asked questions

Official sources

1. AutoArk/EVA-OS on GitHub
2. License: MIT
3. Project website
4. README
5. Releases