Himanshu Gupta

Building video2ctx: video context for AI agents | Open source | Ex-Firecrawl

[Himanshu Gupta @devhims]

we're live on product hunt today 🚀

I ran my own launch video through whisper. "claude code and codex" came out as "plot code and codecs."

that's the problem. most video tools for agents stop at a transcript like that. others process every frame and burn compute.

video2ctx gives your agent only what it needs, from what's said and what's shown, cited to the exact second.

sign up this launch week and get 5,000 free credits. 🎁

would love your support and honest feedback. link in the replies.

[Himanshu Gupta @devhims]

@Joshua_WD Astra high has been super reliable for me. Never felt the need to go any higher.

[Himanshu Gupta @devhims]

@vertigo_labs yeah, totally. thanks so much. would appreciate an upvote. 🙏

[Himanshu Gupta @devhims]

we're live on product hunt today 🚀

I ran my own launch video through whisper. "claude code and codex" came out as "plot code and codecs."

that's the problem. most video tools for agents stop at a transcript like that. others process every frame and burn compute.

video2ctx gives your agent only what it needs, from what's said and what's shown, cited to the exact second.

sign up this launch week and get 5,000 free credits. 🎁

would love your support and honest feedback. link in the replies.

[Himanshu Gupta @devhims]

@Joshua_WD Haha, thanks 😅

[Himanshu Gupta @devhims]

"we" = me, claude code, codex, kiro, and my own agent atlas 😁

[Himanshu Gupta @devhims]

product hunt: https://t.co/uP81huFIPb

[Himanshu Gupta @devhims]

@OpenAIDevs solving video context for agents: https://t.co/x0UWi2oPtF

[阿诚｜AI实战 @byCanen]

用 AI 分析视频，很多工具实际上只拿了字幕。
如果视频主要讲软件操作、图表或产品演示，字幕往往不够。比如讲解者说「点击这里」，Agent 根本不知道这里是哪。
有个开源项目 video2ctx，专门把视频整理成 Agent 能使用的结构化上下文。
除了字幕，还能读取视频元数据、章节、评论、播放列表，以及带时间戳的具体片段。
视觉部分会先建立 Storyboard 索引，只有确实需要看画面时，才用本地 FFmpeg 提取对应时间点的帧，而不是一上来把大量截图全部塞进模型上下文。
安装：
npx skills add devhims/video2ctx
如果只需要研究公开 YouTube 视频，可以使用其中的 youtube-ctx，不必先注册 video2ctx 账号或购买托管 API。
装好后可以给 Agent：
「分析这个 YouTube 产品演示视频。提取完整操作顺序，每一步注明时间戳。遇到字幕无法判断的按钮、菜单和界面状态，再定位对应视频帧。最后输出可复现的操作步骤和证据。」
GitHub：https://t.co/e9ihHfSS7V
目前公开支持的主要视频平台是 YouTube，托管 API 及持续监控属于另一套需要账号的服务。
作者称能明显减少视频上下文 Token，但具体节省比例还需要按自己的视频任务测试。

[Himanshu Gupta @devhims]

@byCanen That's a great summary of video2ctx. Thanks @byCanen!

[Himanshu Gupta @devhims]

install one skill and it works in claude code and codex:

npx skills add devhims/video2ctx

or use the cli, api, or npm package. fully open source ⭐
https://t.co/x0UWi2oPtF

[Himanshu Gupta @devhims]

excited to share what I've been building for the past 11 weeks: video2ctx 🎬

video is still a major blind spot for agents. they either give up, read only the transcript, or sample frames at fixed intervals.

video2ctx gives your agent exactly what it needs, cited to the second, with ~95% fewer tokens. 🤯

fully open source, and launching on product hunt this wednesday, oct 21

[Himanshu Gupta @devhims]

launching wednesday on product hunt. would love your support ❤️
https://t.co/ce8fT1j1Xe

[Himanshu Gupta @devhims]

try it today with 1,000 free credits
https://t.co/WqvFVKTavg

[Himanshu Gupta @devhims]

@TJeparskis check your usage analytics for a clearer picture. try the "by surface" view.

[Himanshu Gupta @devhims]

found the likely cause of the high "web (work)" usage. my dot had started using codex cloud for code reviews without confirming with me. that explains why the weekly limit was dropping so quickly. i've asked it to use its own terminal and filesystem instead.

[Himanshu Gupta @devhims]

Is dot's browser activity counted towards our weekly quota?

appears to have started happening this week. my analytics show "42.8%" web (work), which is highly unusual.

@OpenAIDevs can you clarify? anyone else experiencing this?

[Himanshu Gupta @devhims]

Is dot's browser activity counted towards our weekly quota?

appears to have started happening this week. my analytics show "42.8%" web (work), which is highly unusual.

@OpenAIDevs can you clarify? anyone else experiencing this?

[Himanshu Gupta @devhims]

@thepaulthomson @OpenAI Congrats Paul! I'm building video2ctx, an open-source platform that makes videos queryable for AI agents. Codex has been a huge help with architecture and implementation. Launching next week!

https://t.co/MUazKLWIEf

[Himanshu Gupta @devhims]

Dear Tibo, it would be much better if future resets could be provided as banked resets rather than automatic ones. Anyway, thanks for the reset!

[Tibo @thsottiaux]

o yes… we’re back in action and we’ll reset usage limits for all paid users across codex and ChatGPT work

sorry about the brief disruption!

(and yes we have a special spare codex when things are down to help us out)

[Himanshu Gupta @devhims]

@chetaslua @Simtangara77866 How is the music generated?

[Firecrawl @firecrawl]

We're launching Discord Office Hours - a live space for you to ask questions, share ideas, and connect with our team directly.

Our first session is March 25 at 2pm UTC, and we're throwing in giveaways to kick things off 🔥