Visual-first video documents

The visuals of a video —
not just the transcript.

Every other tool hands your AI a transcript and loses the screen — the diagrams, slides, and code where the meaning actually lives. VidLens reads the screen and returns a structured, AI-ready document your agent can call over MCP.

One moment, two readings

[04:32]“…the request flows from the API gateway through the message queue — as you can see in this diagram.

Transcript onlywords, no screen
“…as you can see in this diagram.”
on-screen diagram — never captured
The referent is lost. Your AI reads a pronoun pointing at nothing.
VidLenswords + what’s on screen
“…as you can see in this diagram.”
On screen [04:32] · architecture diagram
ClientAPI GatewayKafkaWorkerPostgreSQL
The screen, in words your AI can read — cited to the second.
Who it’s for

Made for the people transcripts leave behind.

The same visual layer, delivered three ways — to your agent, to your notes, and to your channel.

Developers & agents

The first MCP server that gives your AI the visuals of a video, not just the transcript.

One line of config and every recording becomes a structured document your agent can call — visual descriptions of what’s on screen, on-screen code, and chapters, all cited to the second.

agent · get_video_doc
› get_video_doc(id)

## On screen [04:32]
Architecture diagram:
Client → API Gateway → Kafka
  → Worker → PostgreSQL
Students & learners

Notes that show the whiteboard, not just the words.

Formulas, diagrams, and slides captured as real text and linked to the second they appeared — so you can skim the lecture instead of scrubbing through it again.

lecture-08-bayes.md
## Key moments
- [12:04] Bayes' rule on the whiteboard

<!-- on screen @ 12:04 -->
P(A|B) = P(B|A)·P(A) / P(B)
Creators

An llms.txt for your channel — make your videos readable by AI search.

Publish clean, structured documents of your own videos at stable URLs, so answer engines and agents can find, read, and cite what you made.

/c/maya-chen/llms.txt
# Maya Chen — video documents

- /v/vector-search   Vector Search in Production
- /v/rag-eval        Evaluating RAG, honestly
- /v/chunking        Chunking is the real model
How it works

Your machine reads. Our server writes.

VidLens is split in two on purpose. The reading happens on your own connection; the document is assembled by our hosted server.

01

Your client reads locally

Point your agent at VidLens. The client runs on your own machine and reads the recording on your connection — never ours.

02

It sends a small payload

Only the transcript and a few deduplicated key frames travel to VidLens. The recording itself stays with you.

03

VidLens assembles the document

Our server describes what’s on screen, structures it into chapters and key moments, and returns one clean Markdown document — callable over MCP.

The video never reaches our servers — only the transcript and key frames do. The reading stays on your machine; we assemble the document.

Anatomy of a VidLens

One recording in. A document your AI can read out.

Summary

A two-sentence TL;DR of the whole recording.

Chapters

The recording split into sections, each with a one-line gist and a timestamp.

On-screen visuals

Diagrams, slides, and charts described in language your AI can read and search.

Timestamped transcript

Every line, speaker-labeled, linked to the second it was said.

On-screen text & code

Slides and code captured as markdown — copy-paste ready.

vector-search-in-production.mdExport .md
# Vector Search in Production
> 18:24 · May 28, 2026 · Maya Chen

## Summary
A pragmatic walkthrough of why naive cosine
search disappoints at scale, and the three
fixes that matter.

## On screen [04:32]
Architecture diagram: Client → API
Gateway → Kafka → Worker → PostgreSQL.

## Transcript
**[03:58] Maya Chen:** If you take one
line home: chunk on structure, not count.

```python
def chunk(doc):
    for s in doc.sections:
        yield s.heading + s.body
```
Built for agents

Your videos, as callable context.

VidLens is an MCP server. Point your agent at it and every recording becomes a structured, AI-ready document it can read directly — visuals included, cited to the second.

See the MCP integration
// mcp.json { "mcpServers": { "vidlens": { "command": "npx", "args": ["-y", "@vidlens/mcp"], "env": { "VIDLENS_API_KEY": "sk-vl-…" } } } }

Give your AI the whole video.

Not just the transcript — the diagrams, slides, and code on screen, as one AI-ready document your agent can call.