NEW Video Agent

Video in. Context out.

Everything your agent needs to understand videosGet started for free1,000 credits · No card required

Try:
Works with:
  • Claude
  • ChatGPT
  • Gemini
  • Cursor

What is video2ctx?

video2ctx makes videos queryable for AI agents and the applications you build. Ask about anything, from spoken words and on-screen visuals to viewer reactions, and every answer comes with timestamped citations you can trace to the original video.

  • Inspect any videoTranscripts, storyboards, frames, and comments as structured data, through skills, CLI, API, or npm package.
  • Ask across videosVideo Agent finds the relevant videos and answers your question across them.
  • Catch what’s newMonitors watch a channel, topic, or search and tell you when something new appears.

Why it matters

Token-efficient
Agents get the answer and the moments behind it, not the whole video. On average, that means about 95% fewer tokens than reading the full transcripts.
Easier on limits and costs
Fewer tokens per question means weekly limits and API budgets go further.
Accurate by design
Answers come from what the video actually says and shows, so claims hold up when you check them.
Fast and multimodal
Speech, visuals, and comments are processed together, and most answers land in under 30 seconds.

Where video comes from

  1. YouTubeFull support
  2. Other platformsNext
  3. Your own uploadsNext

Set it up your way.

Work inside your agent, from the terminal, or in your own code. Each route says when to pick it.

npx skills add devhims/video2ctx

Ask Claude Code, Codex, or any agent that loads skills about a public YouTube video, and it fetches the transcript, comments, and frames itself, right from your machine. No account, API key, or CLI needed.

Browse the skills

Ask a question.
Video Agent watches the video for you.

Video Agent reads transcripts, frames, and comments across up to eight videos, then answers with sources.

Praggnanandhaa vs Vincent Keymer · 14:59

Where did Keymer lose this game?

  1. Metadatatitle, channel, 14:59
  2. Transcript380 segments
  3. Frames4 moments
  4. Comments114 comments
Video AgentAnswered

Pragg was better after Keymer gave up an exchange , but the game was heading for a threefold draw until Qg4 .[1]

The loss came late. Keymer missed draws with Nd3 and Nf3+ , then resigned on move 68 .[1]

Viewers agree: a top comment quotes “Vincent is only human” over the missed Nd3.[1]

1 video reviewed[1] Praggnanandhaa vs Vincent Keymer, agadmator's Chess Channel
Example built from this video’s real transcript and comments.

Monitor a channel.
Know the moment it posts.

Monitors check a channel, topic, or search from hourly to weekly and alert you in the dashboard or by email. Hand any new video straight to Video Agent.

Channel monitorTheo - t3.gg · every hour
  1. 3h agoNo change
  2. 2h agoNo change
  3. 1h agoNo change
  4. NowNew video
  • New video from Theo - t3.ggMeta just caught up with Anthropic
  • New video from Theo - t3.gg · 1 day agoI thought AI was gonna get cheaper?
  • New video from Theo - t3.gg · 3 days agoYou can rewrite typescript in rust for $420k
Meta just caught up with Anthropic · 38:02

What does he claim, and what do the comments push back on?

Video AgentAnswered

Meta’s Muse Wonder Full is the first model outside OpenAI and Anthropic to reach the Fable and Astra tier .[1]

It is fast and token efficient and can orchestrate sub agents , but it is weak at UI .[1]

The top comment (439 likes) asks “How much did they pay you”. He answers on camera that Meta paid nothing .[1]

1 video reviewed · 4 credits[1] Meta just caught up with Anthropic, Theo - t3.gg
Example built from a real monitor, its latest alert, and the answer.

What people use it for.

Three real runs: numbers read straight off the screen, a lecture turned into notes, and three tutorials compared.

How to adjust camera focal length in Blender · ENORMOUS · 1:11

What focal length does the Lens panel show, and when does it change? Use the frames, not the narration.

  1. Storyboard72 tiles, one per second
  2. Frames11 read in full
  3. Transcript30 lines, no numbers
Video AgentAnswered

Before the edit, the Lens panel reads 51.13 mm [1]

After the change in the Camera tab, it holds at 50 mm to [1]

Dragging the viewport handle swings it to 67.75 mm , then back to 49.82 mm [1]

None of these values is spoken. They appear only on screen.[1]

1 video · 11 frames · 6 credits[1] How to adjust camera focal length in Blender, ENORMOUS
Replayed from real runs. Answers are trimmed, not invented.

Start free. Pay when you build.

Every plan includes the dashboard and API. Credits are charged according to the data you request.

Starter

FreeNo card required

1,000 credits once

Explore complete video data in the dashboard and API before committing.

  • Dashboard and API access
  • Transcripts, comments, channels, and playlists
  • Projects and monitoring
  • One credit grant per account
Start free

Builder

Recommended
$20per month

10,000 credits monthly

Build production applications, agents, and internal tools with recurring usage.

  • Everything in Starter
  • Monthly credit renewal
  • Higher project, monitor, and import limits
  • Standard support
Start building

Scale

CustomBuilt around your workload

Custom credit allocation

Run video2ctx inside a product or data pipeline with company-level requirements.

  • Volume pricing and custom limits
  • Invoice billing options
  • Priority onboarding and support
  • Security and SLA options

Tell us what you need at scale.

Share a few details so we can respond with useful limits, pricing, and onboarding options.

By submitting, you agree that we may contact you about this inquiry. See our privacy policy.

FAQs

How far do 1,000 free credits go?

Most reads cost 1 credit, and a fresh search or comments request costs 2. In our tests, a cited Video Agent answer about one video used 4 credits and a three-video comparison used 3, so the free grant covers a few hundred questions. No card is needed, and there is one grant per account.

Do I need an account?

Not to start. The youtube-ctx skill and the all-things-youtube npm package run on your machine with no account or API key. An account adds the hosted API, Video Agent, projects, and monitors.

Which agents does it work with?

Any agent that can load skills or run a command-line tool, such as Claude Code, Codex, Gemini CLI, and Cursor. Install the skills with one command, and add the CLI when you want hosted features.

What can Video Agent read?

Transcripts, storyboards, frames, and comments across up to eight videos per question, each up to two hours long. It uses the storyboard to see the whole video at a glance, then reads full frames for on-screen detail like interface values, slides, and code. Every answer cites the videos it used.

How do monitors work?

A monitor checks a channel, topic, or search on a schedule, from hourly to weekly. The first check records a baseline, and each new video after that raises an alert in the dashboard, or by email once you confirm your address. Starter includes one monitor.

What if a video has no captions or comments?

You are told plainly. Results say when captions are missing or comments are turned off, instead of returning an empty list. Transcripts work on long videos too, and captions can be translated into other languages.

Is it open source?

Yes, under Apache 2.0. The workspace and API are one codebase, so you can use the hosted version, install the package, or run all of it yourself.

Love what you’re building. How can I support it?

That’s really kind of you. I’m building video2ctx on my own, so it means a lot. The Builder plan helps most, as it pays for the servers (and my coffee). Otherwise, a star on GitHub and spreading the word go a long way :)

Liked the preview? Save the full results.

Ask across videos, read what’s on screen, and catch new uploads.
Start with 1,000 free credits.