Research desk and edit bay for faceless YouTube channels

Find the video worth making.

ytr measures channel baselines, outlier videos and search competition. An analyst cites stored evidence for every claim, code checks the citations, and you make the call.1

Get started
simulated
3× baselineOutliers start here
1×Channel baseline

One niche's recent uploads, simulated. Each bar is a video's views against its own channel's baseline, on a log scale. Most land near 1×. The few past 3× are what ytr goes looking for.

ytr answers two questions. Everything else in it exists to answer them with evidence, not with a number that looks precise.

  1. 01

    Which niche should I start a channel in?

    Proven demand, room to compete, real money, and content you can make faceless and original.

  2. 02

    What video should this channel make next?

    What is working in the niche right now, where the gaps are, and what has worked on your own channel.

How it works

Measured by code. Judged by an analyst. Decided by you.

  1. Channel baselines, outlier multipliers, and the age and size of search results come straight from the YouTube Data API. Baselining a channel costs about 2 quota units.3

    measured

  2. An analyst does the judgment work: any agent that speaks MCP reads the method spec, calls the tools and saves a report. Every claim in it points to stored evidence or a URL.2

    measuredextractedinferred

  3. Code checks the citations. A run counts only if a validated report was saved, because the core never trusts the agent's text. Then the verdict is computed from five pass/fail gates.4

  4. Approve, watch or reject. Approving a niche creates a channel profile, and approving a video idea turns it into a brief.

    uv run ytr decide 1 approve

01 Measure simulated
Search results, before anything has been measured.

Gates

Five gates. No score out of 100.

A composite score looks more precise than its inputs are, so ytr uses five pass/fail gates instead. Some are measured, some extracted, some inferred, and code computes the verdict from them.2 Set them yourself: the verdict follows the rule the code runs.

Demand measured

3 or more on-topic outliers (at least 3× baseline, within 180 days, not ad-promoted) across 2 or more small channels with solid baselines: 8 or more videos and a median of 1,000 views. A small breakout channel can count in place of outliers. Long videos and Shorts are judged separately.

Demand

Competition measured + inferred

2 or more core queries each show a gap: at least half the top results are over 12 months old, or 30% or fewer come from channels above 1M subscribers. And no more than 4 small channels already have on-topic outliers.

Competition

Money extracted

2 or more specific offers (affiliate programs or sponsors), each with a URL.

Money

Faceless fit inferred

At least one outlier channel is faceless, and the format doesn't need a presenter. The analyst proposes it and you confirm with watch checks: a video and a yes-or-no question, answered with ytr check.

Faceless fit

Originality inferred

A specific angle that isn't in the top results or in the batch's other titles. You confirm this one with watch checks too.

Originality

Pipeline

Then make it. Six stages from script to live, and nothing is marked live until it is.

  1. 01Script

    The words, kept to one story. Every narration change is logged in the episode's story ledger and checked before it is voiced again.

    script.md · STORY.md

  2. 02Voice

    The narration, voiced from the script and kept in the video's folder.

    voiceover/

  3. 03Build

    An HTML build whose watch page plays the full cut. Review it before anything is rendered.

    <build>/watch/

  4. 04Render

    The long video and its vertical cuts for Shorts, TikTok and Instagram.

    upload/*_LONG_*.mp4

  5. 05Package

    Titles A, B and C, thumbnails, description, chapters, tags, captions and disclosures, plus copy for every platform. Not done until the checker passes.6

    upload.json · UPLOAD.md

  6. 06Live

    Marked only when it's really up. A scheduled post is not an upload.

    uploaded.json

Dashboard

One place for every channel and video. Dark like every editing suite, so thumbnails and cuts read true.

sample dataThis is the Overview page that uv run ytr app opens, drawn in HTML. The app only reads your video folders, except for a field you edit.7

  • Player and preview

    The long cut, the Shorts and the narration play in a tab. So does each build's watch page, with no render.

  • Thumbnail gallery

    Every draft in a grid or side by side. Flag one Lead, 2nd or 3rd, then post it.

  • Upload status

    Each thumbnail, title, caption and Short shows Used or Not used. Agents read the same record.

  • Everything to paste

    A, B and C titles and the copy for YouTube, Shorts, TikTok, Instagram and X, each with Copy and Edit.

  • Research

    Start batches, read reports with links to their evidence, answer watch checks, then approve, watch or reject.

  • Find anything

    Any video, file, build, agent thread or batch. Paste a preview link and it plays in a tab.

Agents

No agent lock-in. Any agent that speaks MCP or can call a CLI can run the analysis.2

An MCP server

The same tools for a chat or a headless run, plus prompts that carry the method specs.8

$ uv run ytr mcp
data
discovermeasure_channelsfind_outlierssearch_profileget_videosquota_status
workspace
get_batchsave_candidatesget_candidateget_evidencesave_reporttrack_channels
method prompts
expand_methodniche_methodtopic_method

Headless runs on your own subscription

The batch runner starts Claude Code with no one at the keyboard. Add another agent CLI with a profile in ytr.toml, not a code change.

[runners.codex]
command = ["codex", "exec", "{prompt}"]
timeout_seconds = 900

Connectors

ytr serve puts the same server on the web behind OAuth 2.1, with setup notes for each of these.

  • Claude
  • ChatGPT
  • Gemini
  • Grok Bot
  • Muse

Agents that don't overwrite each other

Several agent sessions can work in one checkout. A guard refuses a write that lands where another session has uncommitted work, and the dashboard lists which sessions touched a video.9

$ python .claude/hooks/work_claims.py status

Quota and terms

Inside YouTube's API terms by default. Anything that isn't is an optional adapter, switched off.2

100
searches a day. Search is for discovery only. Everything else uses ID-based calls.
about 2
units to baseline a channel, out of 10,000 a day.
24 h
of caching, so a run repeated within a day reuses what it has.
30 days
then API data is refreshed or deleted. Run ytr maintain daily and it's done.

Reports keep their conclusions and show stats expired · refresh where numbers used to be. Nothing is ever filled in with made-up numbers.10

LIVE

Start with one seed.

A market or a niche idea. ytr finds and validates up to 5 niches for each, and every command prints the next one to run.

Niche research, end to end
$ uv run ytr research AI          # a market or a niche idea
$ uv run ytr report 1             # read niche 1's report
$ uv run ytr decide 1 approve     # or reject, or watch
$ uv run ytr profile topics       # next-video research for that niche
$ uv run ytr batch run
$ uv run ytr decide 1 approve     # the top idea becomes a brief
$ uv run ytr brief show

Needs Python 3.13, uv, a YouTube Data API key in .env, and Claude Code or another agent CLI for the analysis steps.
The repository is private, so there's nothing to download from this page.