Agnes AI Review: One Free OpenAI-Compatible API for Text, Image, and Video

One API key covering text, image, and video—free, OpenAI-compatible, and with a China site for domestic access. Agnes AI sounds too good to be true. But spread its 5 models out: chat and coding via agnes-2.5-flash, image generation via image-2.1-flash, video via video-v2.0, all behind one base_url, and the openai SDK runs with two lines changed. This piece reviews it as a comparison: each model's best range, the gap against paid options, and where the free quota actually bites.

Scope first: this is based on the AgnesAI-Labs/AgnesAI-Models official README, agnes-ai.com docs, and public HN discussion, as of 2026-08-10. Star counts for the three repos were verified via GitHub (AgnesAI-Models 2,728★, agnes-video-generator 222★, agnes-ai-skill 58★). Free quotas are subject to the official platform platform.agnes-ai.com and may change; this is a representative comparison, not a personal full benchmark.

One: What Is Agnes AI—A Free OpenAI-Compatible Multimodal Gateway

Agnes AI is an official multimodal API gateway whose pitch is one sentence: one key, one OpenAI-compatible interface, text/image/video all in. For developers that means near-zero migration—the code you wrote against the openai SDK works by swapping base_url to https://apihub.agnes-ai.com/v1 and api_key to Agnes's, nothing else.

Two details matter. First, dual sites: international agnes-ai.com plus a China site at agnes-ai.cn. Developers in mainland China connect directly without a VPN—arguably its biggest edge over most overseas free gateways. Second, broad model coverage: 5 models span text chat, coding agents, image generation/editing, and video generation, not just chat. Docs live at agnes-ai.com/doc/overview, console at platform.agnes-ai.com.

An ecosystem is already growing: jomeswang/agnes-ai-skill (58★, MIT) plugs Agnes into Codex, Claude Code, WorkBuddy, Claude Desktop, Cherry and other agents; lcy362/agnes-video-generator (222★, MIT) is an open-source self-hosted video generator built on Agnes's free models. This site's Agnes free multimodal workflow SOP covers how to chain them; this piece first covers what Agnes offers and whether it's worth using.

Two: Five-Model Comparison—Specs Table

Five models across type, endpoint, capability, and best scenario.

A few notes. First, both text models (2.5-flash and 2.0-flash) have vision-language capability—they can "see images," not just chat. Second, the image models use /v1/images/generations, the same endpoint name as OpenAI DALL-E, so drop-in is clean. Third, the video model uses /v1/videos as an async task API—you submit a task, get a task_id, and poll for the result, not a synchronous response (covered in pitfalls below). Fourth, 2.5-flash is the newest flagship with the fullest capability set; 2.0-flash is lighter, for scenarios that don't need the latest and want to save quota.

Three: One by One—Each Model's Best Range

agnes-2.5-flash: main chat, coding, and vision workhorse, fullest capability. The newest flagship text model, with coding, agent, tool-calling, multi-turn reasoning, and image understanding all in. One model covers "chat + write code + read screenshots to fix bugs + chain tool calls"—Agnes's core selling point. Best for: a personal developer's daily AI assistant, the brain of an agent workflow, any task needing vision.

agnes-2.0-flash: lighter text model, saves quota. The previous flash generation, supporting reasoning/coding/tool-calls/streaming/image understanding/agent—just not as new as 2.5. When you don't need peak capability and want to reserve quota for image and video, run chat on this one. Best for: light chat, intermediate steps in agent orchestration, batch calls that save quota.

agnes-image-2.0-flash: basic text-to-image. Supports text-to-image and image-to-image, with input/output via URL or Base64. Enough for illustrations, covers, quick concept art. Shortcoming: resolution and detail trail 2.1. Best for: quick generation, batch illustrations where quality isn't critical.

agnes-image-2.1-flash: high-quality generation + image editing. The 2.0 upgrade, focused on "high-density visual generation" (denser detail, higher quality), image editing, and flexible sizes. For production-grade images, image edits, or non-standard sizes, pick this. Best for: production-grade generation, image editing, flexible-size scenarios.

agnes-video-v2.0: video generation, async tasks. The only video model of the five, supporting text-to-video, image-to-video, multi-image video, and keyframe animation. Note it's async: submit a task, get a task_id, poll until done. Free-tier video quota is the tightest (see pitfalls). Best for: short video generation, concept videos, video assets for content creation.

Four: Agnes vs Paid and Other Options

A single platform can still be compared—set Agnes in the "free multimodal API" lane against OpenAI's paid official API and other free gateways, to see where it wins and where it lags.

The conclusion is direct: Agnes's edge is the "free + China site + OpenAI-compatible + truly multimodal" four-in-one. OpenAI's official API has the fullest modal coverage but costs money and isn't usable domestically; other free gateways mostly cover text only and lack a China site. Agnes stacks all four, which is especially attractive for developers in China—no VPN, no code changes, one key for text/image/video. Against the paid models in this site's LLM cost-performance comparison, Agnes wins on "zero-cost start"; against another free window, WorkBuddy, Agnes wins on multimodal coverage.

Five: How to Use—OpenAI SDK Drop-in

Migration cost is the key metric for "easy to use." Agnes is fully OpenAI-compatible; with the openai SDK, change two lines:

from openai import OpenAI

client = OpenAI(
    api_key="agnes-xxx",           # your Agnes key
    base_url="https://apihub.agnes-ai.com/v1",  # Agnes base
)

# Text chat (streaming)
stream = client.chat.completions.create(
    model="agnes-2.5-flash",
    messages=[{"role": "user", "content": "Explain MCP in one sentence"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

# Image generation
img = client.images.generate(
    model="agnes-image-2.1-flash",
    prompt="Minimalist cover, AI tools theme",
    n=1,
)
print(img.data[0].url)

Image editing and video generation work the same way, with OpenAI-matching endpoint names. Video goes through /v1/videos as an async submit and needs polling the task status. Full examples are in the official docs at agnes-ai.com/doc/overview.

Six: Selection Advice—Pick by Need

The last row is the ecosystem add-on: to use Agnes inside your own agent, agnes-ai-skill already does the adapter. For fully offline zero-cost video, lcy362/agnes-video-generator is an open-source self-host option, and OpenMontage can also use Agnes as a free model source.

Seven: Three Pitfalls

One: free quota is limited, video quota is tight—don't run it as production. The Free/default tier gives text 20 RPM, image by resolution, and video by RPM+quota—video is the tightest. It's fine for prototyping, personal projects, and content creation; for production traffic, check the tiers and limits on platform.agnes-ai.com first. The free tier won't hold a production load.

Two: video is async—poll it. agnes-video-v2.0 goes through /v1/videos: submit, get a task_id, and poll the task status until done—not a synchronous result. First-timers often write it as a synchronous wait and hang. Follow the official async polling pattern, with proper timeouts and retries.

Three: China site and international site may differ in models/quotas—check official. agnes-ai.cn and agnes-ai.com are two sites; the model list, free quota, and available features may not be fully identical. Developers in China should prefer the China site, but verify actual capability against platform.agnes-ai.com in real time—don't assume the two sides are fully symmetric.

Eight: FAQ

Q: Is Agnes AI really free? Any hidden fees? A: There's a free tier (Free/default)—text 20 RPM, image rate-limited by resolution, video by quota, at no cost. But the free quota is finite; exceeding it or high-frequency commercial use means upgrading to a paid tier—pricing per platform.agnes-ai.com. "Free" means a usable free tier to start, not unlimited free.

Q: Usable in China? Need a VPN? A: No. Agnes has a China site at agnes-ai.cn, directly accessible domestically—its biggest difference from most overseas free gateways. The API base is https://apihub.agnes-ai.com/v1.

Q: Is it identical to the OpenAI official API? A: The interface is compatible (endpoint names and parameter structure match), and the openai SDK works by changing base_url and api_key. But the models are Agnes's own (agnes-2.5-flash etc.), not GPT, with different capability and behavior. The video endpoint /v1/videos is an Agnes extension; OpenAI has no exact synchronous video equivalent.

Q: Can Claude Code / Codex use it directly? A: Yes. jomeswang/agnes-ai-skill (58★, MIT) is an Agent Skill supporting Codex, Claude Code, OpenClaw, Claude Desktop, Hermes, WorkBuddy, and Cherry, wiring Agnes in as the model backend for these agents.

Q: What if the free quota runs out? A: Upgrade to a paid tier (per official), or use lcy362/agnes-video-generator open-source self-hosting for zero-cost video and move high-frequency tasks local. This site's Agnes free multimodal workflow SOP is specifically about building a zero-cost workflow.

References

• AgnesAI-Labs/AgnesAI-Models GitHub repo (2,728★)
• lcy362/agnes-video-generator GitHub repo (222★, MIT)
• jomeswang/agnes-ai-skill GitHub repo (58★, MIT)
• Agnes AI official docs
• Agnes AI China site
• This site: Agnes free multimodal workflow SOP ｜ OpenMontage open-source resource ｜ LLM cost-performance comparison ｜ WorkBuddy free window