Run the analysis that wasn't possible before.
Your agent gets direct, read-only access to the database: channels, videos, transcripts, brand mentions and view snapshots. Ask in plain language. It writes the query, shows its work, and keeps the receipts.
Which channels ran a cookware sponsorship in the last 90 days, over 100K subscribers?
used tl db es '{"size":0,"query":{"term":{"sponsored_brand_mentions":"5612"}},"aggs":{"by_channel":{"terms":{"field":"channel.id"}}}}'
14 channels matched. The top five by subscribers:
| Channel | Reads | Subscribers |
|---|---|---|
| Home Cook Weekly | 4 | 612,000 |
| The Repair Bench | 3 | 488,000 |
| Copper & Cast | 2 | 301,000 |
| Weeknight Kitchen | 2 | 174,000 |
Every read is linked to its video and its timestamp, so you can watch it.
Analysis that wasn't possible
Not faster lookups. Questions that had no answer before, because nobody could hold the transcripts, the view curves and the deal history in the same place at the same time.
Is this audience real?
Your agent pulls the view curve for every recent video, reads the shape of it against 8.5B snapshots of real channels, and tracks the like, comment and view ratios over time. Then it benchmarks all of that against similar channels in the same niche, so "this spike is not organic" means it does not look like the neighbours, not that it looks odd in isolation.
That is the channel authenticity skill: a 0 to 100 score with the red flags ranked, before you reply to the pitch.
Vet this channel before we book it.
used tl snapshots video 5f2a91c --channel 12345 --since 2026-06-01
Authenticity 41 / 100. Three red flags, ranked:
| Signal | vs. peers |
|---|---|
| Views in first 48h | 7.1x |
| Comments per 1K views | 0.2x |
| Likes per 1K views | 0.4x |
The curve front-loads and then flattens. Twelve commenters posted the same phrase.
Find every sponsored read for this brand in outdoor gear, with the clip.
used tl db es '{"query":{"match":{"transcript":"cold brew"}},"highlight":{"fields":{"transcript":{}}}}' --highlight
| Channel | Video | At |
|---|---|---|
| Trailhead Weekly | Three days, one pack | 4:12 |
| Backcountry Bench | Winter kit, rebuilt | 1:48 |
| Ridge & River | What I actually carry | 9:05 |
Each row carries the spoken snippet, so you can read the script and jump to the clip.
Every read a brand has ever bought, in one pass.
Brand mentions are indexed against the transcript, so each one carries the snippet that was actually spoken and the timestamp it was spoken at. Ask for every sponsored read a brand ran across a whole category and your agent comes back with the list, the exact clip and the moment in it.
Sponsored and organic are both there, and they are labelled, so you can also see who talks about the brand without being paid.
Bring your own numbers into the same context.
Your campaign results live in a spreadsheet nobody else has. Your agent can read that file and query the database in the same session, so it is the agent that joins your CPA per channel to our view curves, audience and sponsorship history. Nothing you upload goes anywhere: the join happens on your machine.
That is the question a dashboard cannot answer, because a dashboard has never seen your numbers.
Join q3-results.csv to the channels and tell me what the winners have in common.
used tl db pg "SELECT channel_name, subscribers, projected_views FROM thoughtleaders_channel WHERE id IN (…) LIMIT 500 OFFSET 0" --csv
Read 38 rows from your file, matched 34 to channels. The winners share three things:
- A flat view curve after week two, not a spike.
- Two or more brands that came back for a second read.
- Under 400K subscribers, so the read is cheaper per view.
Why a tool your agent runs, not just another MCP
Most products hand your agent an MCP, which is an API with a nicer name. That shape costs you three things.
| An MCP | A tool your agent runs | |
|---|---|---|
| How the answer arrives | Every row of every answer is poured into the agent's memory, in full. | The agent filters the answer, saves it to a file, and comes back to it when it needs it. |
| What a long session feels like | Context fills up, so the session gets slower and more expensive as it goes. | Context stays small, so it is much cheaper on agent usage and it keeps its train of thought. |
| What you can ask | Only the questions the API author thought of. | Anything the database can answer, including the join nobody planned for. |
There is also an MCP server for agents that only speak MCP. It reaches the same data. We just think the agent does better work when it can keep the answer in a file instead of in its head.
Works where you work
It is a command, so it runs wherever your agent runs. Claude, Codex, Gemini, Grok, whatever your team has open.
- Claude Code
- tl setup claude installs the plugin, and the skills arrive with it.
- Claude Cowork
- Runs in its terminal, same login, same balance.
- Codex
- tl setup codex installs the skills where Codex reads them.
- Gemini CLI
- tl setup gemini installs the skills where Gemini reads them.
- OpenCode
- tl setup opencode installs the skills into OpenCode's skills directory.
- Cursor
- Runs in Cursor's terminal, no extra setup.
Or any agent that can run a command.
Skills: whole analyses your agent already knows
A skill is a job someone has already worked out, step by step. Set your agent up and they arrive with it.
/tl-keyword-research YouTube Keyword Research
Turns a topic into a validated search filter over every YouTube title, description and transcript we index, plus the channels and videos it selects.
What it does →/tl-channel-authenticity Channel Authenticity Check
Vet a YouTube channel for fake views, bot comments, and non-organic engagement before you book a sponsorship. Returns a 0–100 authenticity score with ranked red-flag findings — calibrated on real bought-view and comment-farm investigations.
What it does →/tl-top-partnerships Top-Performing Partnerships
See which of your sponsorships actually performed — ranked by effective CPM once the videos went live, with each deal's live eCPM measured against the projection you bought at. Delivered as a two-tab Google Sheet, with re-book flags for renewal planning.
What it does →/tl-views-guarantee Views Guarantee Calculator
Get the views guarantee to write into the deal, plus your odds of hitting it. Point it at a creator and it simulates 10,000 buys by resampling that channel's real 30-day view counts, then returns the smallest video bundle that clears 80%.
What it does →Build on it
Because the access is at the database level, teams stop asking us for features and just build the thing.
- An internal dashboard for the creators they already work with, refreshed every morning.
- A Sheets add-in, so the workbook the team already lives in pulls fresh deal history.
- A Slack bot that vets a channel the moment someone drops a link in the channel.
The schema is documented and the output is JSON or CSV, so there is nothing to reverse engineer.
What it costs
Try it free: every account gets a weekly allowance, spent across every surface. Add a team seat when you want reports and sharing with the rest of the team. Move a seat to the $200 tier when someone's agent needs a much larger allowance.