试了一圈国产 AI Agent 的免费额度,百度阿里豆包先淘汰了,剩下这 8 个I tested a bunch of domestic AI agent free trials, and Baidu, Alibaba, and Doubao were the first to be eliminated. Here are the remaining 8.


AI free token/AI免费积分

东拼西凑,又是一条好 workflow

cover

最近各家大厂的 Agent 产品扎堆上线,云端沙箱、本地客户端、CLI 工具……形态五花八门,免费额度也各有玄机。

我一开始想做的很简单:评一评网页上直接能用的 Agent。不用下载,不用配置环境,上班打开一个浏览器标签页就开干——对普通人来说,IDE 和 CLI 那套东西太重了。

但评着评着发现:国内真能干活的网页版 Agent 确实没那么多,而有两个非网页端的产品(ZCode 和 Kimi Work)免费额度实在太香,不放进来对不住大家。

所以这篇的结构是:先讲淘汰的 → 再讲 6 个网页端推荐 → 最后单独聊 2 个非网页端的羊毛选手。我会标清形态,你们自己判断要不要试。

💡 非广告,纯野生用户视角。


先说被我淘汰的

百度千帆 Agent —— 听不懂人话。拿到指令就开始续写散文,完全不管需求。试了两三轮,放弃。

阿里 Qwen3.7Max Agent —— 三个任务挂了三个:网络超时、说完成不了、生成的网页一眼流水线制品。没有继续试的义务。

豆包办公任务模式 —— 给了指令一下就 bug,没耐心。

不是我挑剔。如果公司做的是给普通人用的 Agent,第一印象挂了就是挂了。没有人会花半小时适应一个工具——尤其后面还有活儿赶着。


🌐 6 个网页端,打开浏览器就能用

1. 字节 TRAE Work

一句话: 体验非常完整的网页端 Agent,从代码工具进化成了全能协作平台。

● 免费额度: 全功能开放,2 个云端任务并发,高峰排队

● 模型: Doubao-Seed-2.1 Pro/turbo、MiniMax-M3、GLM-5.2、DeepSeek-V4-Pro、Kimi-K2.7-Code。

● 亮点: Work / Code / Design 三模式,联网 Ubuntu 沙箱,支持 MCP 扩展、Skills 技能系统(含 Figma 转代码/前端设计等)、cron 定时任务、飞书深度集成(云文档/多维表格/日历/消息一键授权)、GitHub 集成(创建/检查 PR),全局+项目两级记忆

● 链接: https://work.trae.cn/

TRAE 一开始更像个 VS Code 形态的代码工具,后来逐步走向 cowork 场景,现在基本全能了。

2. 腾讯 WorkBuddy

一句话: 腾讯生态的全能沙盒,集成做得扎实。

● 免费额度: 每月 500 Credits(6.30 前注册首月赠 2000 Credits),但复杂任务一个就能烧 1000

● 模型: DeepSeek-V4 系列、GLM-5.2、MiniMax-M3、Kimi-K2.7

● 亮点: 联网 Ubuntu 沙盒,深度集成腾讯生态,还支持 GitHub、Notion 等外部服务接入。日常办公场景很顺手

● 槽点: 积分消耗比预期快

● 链接: workbuddy.cn/app

图片

我用下来,网页端真正好用的就是 TRAE 和 WorkBuddy。一个背靠字节/飞书,一个背靠腾讯生态,各自把集成做深了,体验就是不一样。

3. 字节 扣子 Coze

一句话: 零代码搭建智能体,偏娱乐和内容创作场景。

● 免费额度: 每日 1500 积分,约 2-3 个简单任务

● 模型: GLM-5.1、doubaoseed-2.1pro/turbo、minimaxm3、DeepSeek-V4 等可选

● 亮点: 拖拽式工作流编排,插件/技能生态完整,长期记忆 + 用户画像。和 Seedance 集成后做视频那条龙服务很顺,文本创作、生活助理、生成应用也都能搞

● 槽点: 积分用太快

● 链接: coze.cn

Coze 和 TRAE 虽然都是字节的,但用起来场景感完全不同。TRAE 是「帮你干活的同事」,Coze 更像「帮你搭玩具/做内容的工具箱」。

4. 智谱 AgentMore

一句话: 智谱自家的 GLM-5.2 Agent,技能多但集成少。

● 免费额度: 每天 200 积分,约 3-4 个简单任务

● 模型: GLM-5.2/5.1(仅自家模型)

● 亮点: PDF/PPT/Excel/图表/网页全格式交付,定时任务 + 心跳轮询,最多 5 个 AI 角色群聊协作,长期记忆 + 用户画像

● 链接: agentmore.chatglm.cn

5. 小米 MiMo Claw

一句话: 金山办公深度集成,文档闭环好,但产品稳定性存疑。

● 免费额度: 每日 4 小时会话 + 千次工具调用

● 模型: MiMo-V2.5-Pro(仅自家模型)

● 亮点: Word/Excel/PPT/PDF 全格式不出页面处理

● 槽点: Free 用户无长期记忆,4 小时后 session 关闭。而且我想回去接着体验的时候,之前的 session 全丢了,完全不知道怎么重新开始 😂

● 链接: mimo.mi.com

6. MiniMax

一句话: M3 模型的官方 Agent 入口,能力在但集成薄。

● 免费额度: 谜。亲测每天大概能处理 2 个任务,积分耗完即停

● 模型: MiniMax-M3(仅自家模型)

● 亮点: 代码文档多模态,长程任务处理

● 槽点: 额度规则不明确

● 链接: agent.minimaxi.com


🐑 2 个非网页端,但羊毛太香不得不放

7. 智谱 ZCode(CLI 工具)

一句话: 国产 Codex,纯粹的代码重构利器。

● 免费额度: 注册后连续 5 天,每日 300 万 Tokens GLM-5.2 + 200 万 Tokens GLM-5-Turbo。不知活动何时结束

● 形态: CLI,需在 ZCode 里登录 z.ai 账号(bigmodel 账号不行)

● 亮点: 代码能力极强,目前正在帮我全权重构整个游戏项目,真有 Codex 那味儿

● 槽点: 高峰期速度稍慢。写代码太投入会直接切回英文思考。说实话我已经完全把它定位成代码工具了,想不起来用它办公

● 链接: zcode.com

图片

8. 月之暗面 Kimi Work(桌面客户端)

一句话: WebBridge 控制真实浏览器是杀手锏,但整体还在早期。

● 免费额度: 公测阶段全功能开放,额度不透明。6.30 前桌面端处理任务额度消耗仅为网页端一半

● 形态: 本地客户端,chat + work 双入口

● 模型: Kimi-2.6-Agent

● 亮点: WebBridge 用你自己的登录态和 Cookie 操控浏览器,长短记忆 + MCP 工具扩展

● 槽点: 还没觉得非常智能,像是给 Chrome 包了个壳就出来了的初级形态。用的人多速度也慢

● 链接: kimi.com

图片


体感分层:这 8 个到底差在哪

用了一圈下来,我的感受是它们明显分成三个梯队:

第一梯队:打通了生态的全能选手  TRAE Work、WorkBuddy。不只是模型强,而是真的把飞书/腾讯/GitHub/Notion 这些工具串起来了,你能感觉到它理解你的工作流,给的回答是「满血」的。

第二梯队:单项突出,场景明确  ZCode(纯代码神器)、Coze(内容创作 + 工作流编排)、Kimi Work(浏览器操控独一份)。各有各的杀手锏,但不是「什么都能干」的类型。

第三梯队:模型能力在线,但产品还没打通  AgentMore、MiMo Claw、MiniMax。只用自家模型,没有接入外部生态,还停留在发挥模型本身的长程任务能力阶段。不是不能用,但缺少那种「被工具串联起来」的顺滑感。


同一个模型,不同平台判若两人

额外说个有意思的发现:同一个 GLM-5.2,在不同 Agent 里性格完全不同——

● WorkBuddy:聪明不讨好,爱干活

● TRAE:简洁直给,也爱干活

● AgentMore(智谱自家):更贴近出厂状态,谦逊不张扬

● ZCode(智谱自家):代码狂魔,写嗨了直接切英文

大概是内置提示词调教的差异。同一个底座模型,包装不同体验就是不同。


怎么搭着用?一张表说清楚

● 写代码 / 重构项目:ZCode、TRAE Work — 额度大,模型强

● 日常办公文档:WorkBuddy、MiMo Claw — 沙盒全能 / 金山办公深度集成

● 操控真实浏览器:Kimi Work — 目前独一份

● 内容创作 / 做视频:Coze — Seedance 集成,一条龙

● 多智能体脑暴:AgentMore — 群聊 + 定时推送

别只盯着一个平台薅,根据任务类型组合着用。白嫖有道,拼着用才是正经。

免责声明: 所有免费额度信息基于 2026 年 6 月 24 日实测,平台政策可能随时调整,请以官方最新规则为准。

图片

作者lili & Opus地点杭州时间2026 夏

 

Patching together, another good workflow was formed.

cover

● 模型: Doubao-Seed-2.1、GLM-5.2、MiniMax-M3、Kimi-K2.7

At first, I simply wanted to evaluate agents that are directly usable on the web. No downloads, no environment setup — just open a browser tab at work and start working. For ordinary people, the IDE and CLI suite is too heavy.

But as I tested, I found that there are indeed not many truly capable web-based agents in China, while two non-web products (ZCode and Kimi Work) offer such generous free quotas that I couldn’t justify leaving them out.

● 链接: coze.cn

图片

4. NVIDIA Ising-Calibration-1.5

Baidu Qianfan Agent — Doesn’t understand human instructions. It starts rambling prose as soon as it gets a command, completely ignoring the requirement. Tried two or three rounds, then gave up.

● 免费额度: 完全免费

Doubao Work Task Mode — Bugged immediately after receiving a command, no patience for troubleshooting.

It’s not me being picky. If a company is making an agent for ordinary people, the first impression matters. Nobody will spend half an hour adapting to a tool — especially when there’s work to finish.

🌐 6 web-based agents, usable by opening a browser

● 链接: https://quantum-calibration.nvidia.com/

图片

5. NVIDIA Ising-Calibration-1.5 on Hugging Face

● Models: Doubao-Seed-2.1 Pro/turbo, MiniMax-M3, GLM-5.2, DeepSeek-V4-Pro, Kimi-K2.7-Code.

● 免费额度: 完全免费

● 模型: 也是 NVIDIA Ising-Calibration-1.5

TRAE originally looked like a VS Code-style code tool but gradually evolved into a full-blown coworking platform — now it’s almost all-in-one.

● 链接: https://huggingface.co/nvidia/Ising-Calibration-1.5

图片

6. NVIDIA Ising-Calibration-1.5 on NVIDIA Ising-Calibration

● Models: DeepSeek-V4 series, GLM-5.2, MiniMax-M3, Kimi-K2.7

● 免费额度: 完全免费

● 模型: 还是 NVIDIA Ising-Calibration-1.5

● Link: workbuddy.cn/app

图片

3. ByteDance Coze

Zero-code agent setup, geared toward entertainment and content creation.

Free quota: 1500 points per day, sufficient for 2-3 simple tasks.

Models: Supports GLM-5.1, doubaoseed-2.1pro/turbo, minimaxm3, DeepSeek-V4, and more.

Highlights: Drag-and-drop workflow orchestration, comprehensive plugin/skill ecosystem, long-term memory, and user profiling. Seamlessly integrates with Seedance for end-to-end video production, and also capable of text creation, life assistance, and application generation.

Downside: Points deplete rapidly during use.

Link: coze.cn

Although both Coze and TRAE are from ByteDance, their user experiences are distinct. TRAE functions as a "colleague to handle your work," while Coze is more like a "toolbox for building toys/creating content."

4. Zhipu AgentMore

GLM-5.2 agent by Zhipu, offering diverse skills but limited integration.

Free quota: 200 points daily, roughly 3-4 simple tasks.

Model: Exclusive use of GLM-5.2/5.1.

Highlights: Full format delivery (PDF, PPT, Excel, charts, web), scheduled tasks with heartbeat polling, supports up to 5 AI role collaborations in a chat, and includes long-term memory with user profiles.

Link: agentmore.chatglm.cn

5. Xiaomi MiMo Claw

Deep integration with WPS Office, good for document closure, but product stability is questionable.

Free quota: 4 hours of session time plus thousands of tool calls daily.

Model: Exclusive use of MiMo-V2.5-Pro.

Highlights: In-page processing for Word, Excel, PPT, and PDF formats.

Downside: No long-term memory for free users; sessions close after 4 hours. Returning users find all previous data lost, making it difficult to resume work. 😂

Link: mimo.mi.com

6. MiniMax

Official agent entry for the M3 model, functional but poorly integrated.

Free quota: Unclear, but practical testing shows enough for about 2 tasks before limits are reached.

Model: Exclusive use of MiniMax-M3.

Highlights: Multimodal handling of code and documents, capable of long-range task processing.

Downside: Ambiguous quota rules.

Link: agent.minimaxi.com

🐑 Two non-web-based tools with exceptional value

7. Zhipu ZCode (CLI Tool)

Domestic counterpart to Codex, a pure code refactoring utility.

● Free quota: After registration, 3 million tokens of GLM-5.2 and 2 million tokens of GLM-5-Turbo are available per day for 5 consecutive days. It is unclear when the event ends.

● Form: CLI, requires logging in to ZCode with a z.ai account (bigmodel accounts do not work).

● Highlights: Extremely strong coding ability, currently helping me fully refactor the entire game project, with a feel akin to Codex.

● Downsides: Speed is slightly slower during peak times. When fully immersed in writing code, it will directly switch to thinking in English. To be honest, I have completely positioned it as a coding tool and don't think to use it for regular office work.

● Link: zcode.com

图片

8. Moonshot Kimi Work (Desktop Client)

One-liner: WebBridge controlling the real browser is the killer feature, but the overall product is still in its early stages.

● Free quota: Full functionality is open during the public beta, with quotas being opaque. Before June 30, tasks processed on the desktop client consume only half the quota of the web client.

● Form: Local client, with both chat and work entry points.

● Model: Kimi-2.6-Agent

● Highlights: WebBridge utilizes your own login state and cookies to manipulate the browser, with long and short memory capabilities and MCP tool extensions.

● Downsides: It doesn't yet feel particularly intelligent; it appears to be a basic iteration where a shell was simply wrapped around Chrome. Speed is slow when many people are using it.

● Link: kimi.com

图片

Perception Tiers: Where Do These 8 Differ?

After trying these out, I feel they clearly fall into three tiers:

First Tier: All-around performers who bridge ecosystems — TRAE Work and WorkBuddy. These aren't just about strong models; they actually interconnect tools like Feishu, Tencent, GitHub, and Notion. You can sense that they understand your workflow and provide "full-strength" responses.

Second Tier: Specialized and scenario-specific — ZCode (pure coding), Coze (content creation + workflow orchestration), and Kimi Work (unique browser control). Each has a distinct killer feature, but none are "do-it-all" solutions.

Third Tier: Model capability exists, but product integration is lacking — AgentMore, MiMo Claw, MiniMax. They only use their own models without connecting to external ecosystems, remaining at the stage of simply leveraging the model's long-range capabilities. These are usable but lack the smooth integration of being "linked by tools."

Same Model, Different Personalities on Different Platforms

An interesting observation: The same GLM-5.2 exhibits completely different characters on different platforms —

● WorkBuddy: Smart but reserved, loves to work.

● TRAE: Concise and direct, also loves to work.

● AgentMore (in-house Zhipu): Closer to the factory state, modest and unassuming.

● ZCode (in-house Zhipu): Coding maniac, switches to English when in the zone.

This is likely due to differences in the internal prompt tuning. The same base model results in vastly different user experiences depending on the packaging.

How to Use Them Together? A Handy Table

● Writing Code/Refactoring Projects: ZCode, TRAE Work — Large quotas, strong models.

● Daily Office Documents: WorkBuddy, MiMo Claw — Sandbox versatility/Deep integration with Kingsoft Office.

● Operate real browsers: Kimi Work — currently unique

● Content creation/video production: Coze — integrated with Seedance, all-in-one

● Multi-agent brainstorming: AgentMore — group chat + scheduled push

Don’t just rely on one platform, use them in combination based on your tasks. Smart use is the key to maximizing free services.

Disclaimer: All free quota information is based on testing as of June 24, 2026. Platform policies may change at any time, please refer to official rules for the latest status.

图片

Author: lili & Opus Location: Hangzhou Time: Summer 2026

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry