👀 Visiting this README?

If Subtext clicks with you, please [star the repo ⭐].

It’s the simplest way to show it landed and helps track real interest.

Subtext

Turn YouTube talks into intelligence. We analyze what's said — and what's not.

MIT License MCP Compatible npm

---

Install: npx skills add samanvaya5/subtext

Works with Claude Code, Cursor, Codex, Gemini, Cline, Windsurf, OpenCode, and 20+ other agents.

---

What it does

Feed it YouTube videos. Get strategic intelligence — hidden signals, what speakers don't say, cross-talk patterns, blind spots. Not summaries.

How it works

Discovery ──> Ingest ──> Schema Negotiation ──> Batch Analysis ──> Synthesis ──> Application
(You pick)    (Bulk)     (YOU approve)         (3-5 at a time)   (Cross-talk)  (Your call)

Every stage has a checkpoint. The AI stops. You review. You redirect. Then it continues.

The key innovation is Schema Negotiation (Stage 3): the AI proposes an analysis framework, you modify it until it's right, and only then does analysis begin. This is not a prompt — it's a negotiation.

Install

# 1. Install the skill
npx skills add samanvaya5/subtext

# 2. Say "use subtext" to your agent
#    It will detect the MCP server isn't connected and walk you through setup
#    One script: detects your agent, asks for API key, writes config. No JSON editing.

Requires: Node.js 18+, a YouTube Data API key (free).

Manual MCP server setup

If you prefer to configure manually, add this to your agent's MCP config:

{
  "mcpServers": {
    "subtext": {
      "command": "npx",
      "args": ["-y", "@rethink-paradigms/subtext-mcp"],
      "env": { "YOUTUBE_API_KEY": "your-key" }
    }
  }
}

Or install from source: git clone https://github.com/samanvaya5/subtext.git && cd subtext && ./install.sh

Analysis modes

Mode	What it extracts
Strategic Deep Dive (default)	Hidden signals, power dynamics, omissions, framing
Technical Extraction	Architecture, stack, benchmarks, replicable patterns
Market Intelligence	Positioning, competitive signals, market gaps
Sentiment & Trend	Trending up/down, buzzwords, consensus vs contrarian
Idea Evaluator	Pressure-test your idea against the corpus
Custom	You define it during Schema Negotiation

How it compares

	Batch	Interactive	Hidden Signals	You control the schema
ChatGPT + URL	-	-	-	-
GPT Researcher	Yes	-	-	-
Manual AI chat	Yes	Sort of	If you're skilled	-
Subtext	Yes	6 checkpoints	Default	Yes

Under the hood

Two things in one package:

• The Skill (skill/SKILL.md) — teaches your agent the 6-stage workflow. Checkpoints, schemas, synthesis patterns — all defined.
• The MCP Server (@rethink-paradigms/subtext-mcp on npm) — 14 YouTube tools: search, transcripts (single/bulk), channel videos, playlists, comments, activities, subscriptions, categories.

The skill tells the agent what to do. The MCP server gives it the tools to do it.

Example

You: Research how AI founders talk about regulation. Last 30 days, these 3 channels.

Agent: [Stage 1] Found 14 videos. Here are the titles. Remove any?

You: Drop the product launches. Keep interviews and panels.

Agent: [Stage 2] 9 transcripts fetched. 2 are low-signal Q&A. Skip them?

You: Yes. What schema are you proposing?

Agent: [Stage 3] Proposed framework:
       - Regulatory positioning (avoidance vs engagement)
       - Competitive moat language
       - Fear appeal patterns
       - What they DON'T mention

You: Add "coalition language" — who they frame as allies vs opponents.

Agent: Updated. Consensus reached. Starting analysis.
       ...
       [Stage 5] Cross-talk pattern: 7 of 9 speakers avoid mentioning EU
       regulation entirely while focusing on US policy. Likely a coordinated
       framing choice across channels.

You: Build an opportunity map for a regulatory compliance startup.

Agent: [Stage 6] Done. Saved to research/analysis/opportunity-map.md

License

MIT