# AgnesAI-Models: an OpenAI-compatible gateway wrapped around a five model catalog

> The official gateway repository for Agnes AI, holding the model catalog, quota tables, bilingual error docs and runnable curl, Python and Node examples.

**AgnesAI-Labs/AgnesAI-Models** — Official Agnes AI gateway and model catalog for OpenAI-compatible text, image, video, and agent workflows.

- Repository: https://github.com/AgnesAI-Labs/AgnesAI-Models
- Website: https://agnes-ai.com/
- Stars: 5,219 · Forks: 353
- Language: Unknown
- License: not declared
- Published: 2026-10-06 · Updated: 2026-10-06 · Language: en
- Canonical page: https://hysenlabs.com/projects/agnesai-labs-agnesai-models

## What actually lives in this repository

The tree is short and almost entirely Markdown. At the top level there is a README in English and a Chinese translation, a MODEL_CATALOG.md, a CHANGELOG.md, a SUPPORT.md, a pyproject.toml, a requirements.txt, and two directories, docs and examples. There is no src directory, no Dockerfile and no deployment manifest, which tells you before you read a word that this is documentation for someone else's running service rather than software you can host.

That framing changes what you can judge here. Questions about training, serving, scaling or cost control are answered on the official site, not here. Questions about which model exists, which endpoint it uses, what your quota is and what an error code means are answered here, and answered unusually well.

The repository has 5,127 stars and 348 forks, which is a lot of attention for a docs site, and 81 open issues, which is a lot of issues for the same. Topics list agnes-ai, ai-api, free-api and multimodal-ai, so the audience is developers hunting for a free or cheap multimodal endpoint. The last push was on 2026-08-18 and the repository is not archived.

## Five models across text, image and video

The catalog section of the README is the substance of the project. Five models are listed, each with a type, an endpoint and a one line description of what it is for. Two are text and vision-language models on the chat completions endpoint, `agnes-2.5-flash` and `agnes-2.0-flash`. Two are image models on the images generations endpoint, `agnes-image-2.0-flash` and `agnes-image-2.1-flash`. One is a video model, `agnes-video-v2.0`, on its own videos endpoint.

The descriptions are specific in a way that helps you choose. The 2.5 flash model is called out for upgraded coding, agent workflows, tool calling, multi turn dialogue, reasoning and image understanding. The 2.0 flash entry lists the same capabilities minus the upgrade claim. For images, the 2.1 model is the dense one with flexible sizes and editing, the 2.0 is the baseline with text to image, image to image, and output as either a URL or Base64. The video model covers text to video, image to video, multi image video, keyframe animation, and an async task API.

Two claims in the surrounding prose deserve a skeptic's attention. The About section says the models are ranked on PinchBench, which is a benchmark you will not find a methodology page for in this repository. It also says the suite is purpose built to power agentic tools such as OpenClaw and Hermes, which is a positioning statement rather than a test. Treat both as marketing until you have measured the models on your own workloads.

## Getting a client running against the gateway

Because the API is OpenAI-compatible, the shortest path is to install the OpenAI Python package and point its base URL at the gateway. The README's quick start is two shell lines and then a streaming call.

```bash
pip install -r requirements.txt
export AGNES_API_KEY="your_api_key_here"
```

```python
from openai import OpenAI

client = OpenAI(
    api_key="YOUR_AGNES_API_KEY",
    base_url="https://apihub.agnes-ai.com/v1",
)
```

The important line is the base URL. Everything else is stock OpenAI SDK usage, which is the practical argument for compatibility: your existing integration changes one constant and a model string rather than a client rewrite. The base URL is also listed in the quick links table as https://apihub.agnes-ai.com/v1.

The dependencies behind that install are minimal, declared in requirements.txt as openai at 1.40.0 or newer and requests at 2.32.0 or newer. The pyproject.toml names the package agnes-ai-examples at version 0.1.0, requires Python 3.9 or newer, and configures ruff with a line length of 88 targeting py39. Example clients exist in three languages, with curl, Python and Node.js directories under examples.

## Rate limits and quotas written down as numbers

Most API documentation buries the limits, or states them in marketing language. This README does the opposite and prints two tables with a date on them, and the date is the part worth reading.

The user plan table gives requests per minute for text, image and video. Free and default is 20 actual RPM for text models and 1 actual RPM for video. Enterprise is 40 actual RPM for text and 2 for video. Token Plan is 1,000 actual RPM for text and 5 actual RPM for video, plus 500 seconds of video per day. Image RPM is not a flat number, it is resolution specific, and higher plans raise the 1K and 2K limits.

The subscription quota table is a different axis, usage included with a paid plan. Starter gets 1,500 requests per five hours and 15,000 per week for the 2.0 flash text model. Plus is five times that at 7,500 and 75,000. Pro is thirty thousand per five hours and three hundred thousand per week. Image plans are flat at 4,000 images per day across all three tiers, and video is flat at 500 seconds per day.

The caveat sits directly above the tables: these are current public reference values as of 2026-06-28, with base Token Plan quotas published 2026-06-22 and video RPM limits updated 2026-06-28, and they are described as operational limits rather than permanent guarantees. Read that as the project does, confirm anything load bearing in the platform console before you ship.

## Two sites, one documentation version

Agnes AI runs an international site at agnes-ai.com and a China site at agnes-ai.cn, and the README says to pick the one that matches your region and service needs. Developer documentation lives at agnes-ai.com/doc/overview, the console is at platform.agnes-ai.com, and the API base URL is hosted at apihub.agnes-ai.com/v1.

The repository tracks itself with a documentation status block rather than a version field in a manifest. Public documentation version is 2026.07.30, last updated 2026-07-30 in Asia/Singapore, source of truth is named as the official website and API platform, and the change notice repeats that model availability, rate limits, pricing and quota rules can change at any time.

That block is unusual and honest in equal measure. It tells you the README is a mirror rather than the authority, and that the changelog in this repository records public documentation, model, quota and integration updates rather than code releases. There is no releases section and no tagged version history, so you cannot pin to a known good state of the API contract through this repository. For a gateway that promises compatibility with the OpenAI surface, that absence is a real difference from vendors who publish dated client libraries.

## Where the docs hand off to the platform

The developer resources table is a map of the repository's own division of labour, and it is worth reading as a statement about what this project does not do.

MODEL_CATALOG.md holds per model endpoints, current reference limits and compatibility notes. CHANGELOG.md holds public documentation, model, quota and integration updates. SUPPORT.md says where to ask for help and what belongs in issues versus discussions. Under docs, TROUBLESHOOTING.md covers API error codes, a debugging checklist and retry guidance, ERROR_CODES.md is a bilingual table of common API status codes with causes and recommended fixes, FAQ.md covers access, limits, models and video polling, TOKEN_PLAN_FAQ.md covers Token Plan access types, RPM limits, subscription quotas and API key limit pools, and DISCUSSIONS.md lists recommended categories.

So the boundary is clear. Authentication, key management, billing and any per account decision live on the platform, not here. What you can do offline, from a clone, is learn the model list, learn the error codes, and run the examples against a key you already hold. For an API catalogue that is the right division, and the parts it does own are unusually complete for a repository with no code in it.

## Conclusion

This repository is a front door rather than a codebase. There is no server here, no training code and no infrastructure definition, just documentation and example clients for a hosted API at apihub.agnes-ai.com/v1, and it is honest about that. What makes it worth reading is that the limits are published as numbers with dates attached, the OpenAI SDK works against it with only a base URL change, and the error code guide is bilingual. The traps are equally clear. There are no releases, so there is no version history to pin against, and the quota figures are operational values from late June 2026 that the README itself says can move. Start at MODEL_CATALOG.md if you are choosing a model, at docs/ERROR_CODES.md if your first request failed, and at the examples directory if you want to see the smallest possible working client before writing your own.

## FAQ

### Who is Agnes AI?

Agnes AI describes itself as a frontier AI company that trains text, image, video and reasoning models in house, and publishes them through an OpenAI-compatible gateway. The company runs an international site at agnes-ai.com and a China site at agnes-ai.cn.

### Does Agnes AI free?

There is a free tier with published limits: 20 actual requests per minute for text models and 1 actual request per minute for video, with resolution specific image limits on top. Paid plans raise text RPM to 40 on Enterprise and 1,000 on Token Plan, and the README notes these are reference values as of 2026-06-28 rather than permanent guarantees.

### Which Agnes AI models are available and what are their endpoints?

Five models are listed. agnes-2.5-flash and agnes-2.0-flash are text and vision-language models on /v1/chat/completions, agnes-image-2.0-flash and agnes-image-2.1-flash serve generation and editing on /v1/images/generations, and agnes-video-v2.0 handles text to video, image to video and keyframes on /v1/videos through an async task API.

### How do I point the OpenAI SDK at the Agnes AI API?

Install openai and requests, export your key as AGNES_API_KEY, then construct the client with base_url set to https://apihub.agnes-ai.com/v1. The README's Python quick start does exactly this with a streaming call to agnes-2.5-flash, and worked examples live under examples/python, examples/curl and examples/node.

## Sources

- [AgnesAI-Labs/AgnesAI-Models on GitHub](https://github.com/AgnesAI-Labs/AgnesAI-Models)
- [Issues](https://github.com/AgnesAI-Labs/AgnesAI-Models/issues)
- [Project website](https://agnes-ai.com/)
- [README](https://github.com/AgnesAI-Labs/AgnesAI-Models/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/agnesai-labs-agnesai-models