Skip to main content
Glama
Channel Brains

Channel Brains

by Pu11en

Channel Brains

Tell your AI to install it. It does.

CI Release Python License: MIT

Channel Brains is a local MCP server that indexes the public captions of any YouTube channel into a searchable database on your own machine — then lets your AI coding agent query it with timestamped citations back to the exact second of video.

No API key. No cloud. No account. It runs entirely on your computer.

Using ChatGPT instead? See Also on ChatGPT.

Paste this into your local AI coding agent (Codex, Claude Code, ZCode, or Hermes):

Get this on yourself: https://github.com/Pu11en/channel-brains

That's the whole instruction. The agent identifies its own client, installs the right plugin, and verifies the six tools. You don't need to know what MCP or a plugin marketplace is.

Why

Every podcast, lecture, and dev stream is a library of knowledge — but video is unsearchable. Channel Brains turns a channel's spoken content into a database your agent can cite. Ask a question, get the answer with a link to the exact moment someone said it.

Related MCP server: tubescribe-mcp

See it

Point Channel Brains at a channel, wait for it to index, then search. Here it is searching Matt Pocock's channel for "generics" — note the timestamped citation back to the source video:

$ channel-brains search_brain --brain-id 8c45c891e781 --query "generics"

rank 1 · Generics: The most intimidating TypeScript feature
       https://youtu.be/dLPgQRbVquo?t=0  (0:00)
       "…we are going to be focusing on 10 tips to make you a master of
       typescript generics… they give you the power to make abstractions, to make
       your code a lot more DRY…"

rank 2 · A Complete Guide To Vercel's AI SDK
       https://youtu.be/mojZpktAiYQ?t=965  (16:05)
       "…use generateObject instead of generateText… we pass it a schema with a
       Zod schema…"

rank 3 · Generics: The most intimidating TypeScript feature
       https://youtu.be/dLPgQRbVquo?t=45  (0:45)

Every result links to the precise second the words were spoken. Your agent gets grounded, quotable evidence — not a hallucination.

How it works

  1. You give your agent a channel URL.

  2. Channel Brains ingests the public captions via yt-dlp (no API key), selecting up to 50 videos by view count, and stores them in a local SQLite FTS5 database.

  3. Ingestion is resumable and rate-limit-aware — it pauses cleanly on HTTP 429 and picks up where it left off.

  4. Your agent searches that database and returns matches with timestamped YouTube links.

The local version is deliberately small:

  • No hosted service, Docker, browser dashboard, embeddings, or LLM runtime

  • No YouTube API key, database server, or cloud account

  • One local SQLite database per user, protected by a cross-process ingestion lock

  • One channel ingestion job at a time, with resumable per-video progress

  • The only network requests are public YouTube requests made by yt-dlp during ingestion

Install

The easy way — paste this to your local AI coding agent and let it install itself:

Get this on yourself: https://github.com/Pu11en/channel-brains

The manual way — requires uv. MCP clients launch the pinned production release with:

uvx --from "git+https://github.com/Pu11en/channel-brains@v0.1.4" channel-brains-mcp

Verify the installation without starting the MCP server or contacting YouTube:

uvx --from "git+https://github.com/Pu11en/channel-brains@v0.1.4" channel-brains-mcp --check

A successful check prints one JSON object with "status": "ok", "transport": "stdio", and "tool_count": 6.

Locally, the server communicates only through standard input and output. Do not run it as an HTTP service yourself. HTTP mode exists only for the hosted ChatGPT version.

Note: Installation requires a local AI coding agent (Codex, Claude Code, ZCode, or Hermes) that can run commands on your computer. A web-only chat (ChatGPT, Claude.ai) cannot install a local MCP server. ChatGPT users can use the hosted version below.

Also on ChatGPT

A hosted version of Channel Brains is available as a ChatGPT plugin. It is pending review in OpenAI's plugin directory.

The six tools

Channel Brains exposes exactly these six tools to your agent:

Tool

Purpose

create_brain

Validate a channel URL, persist a brain, and start local ingestion.

get_brain_status

Read a snapshot or wait locally for completion, with progress and no YouTube requests.

list_brain_videos

Page through indexed, skipped, pending, and failed video records.

search_brain

Search local FTS5 caption chunks and return timestamped YouTube citations.

get_video_transcript

Page through stored caption chunks for one indexed video.

delete_brain

Remove a completed brain and all its local records. Active ingestion is refused.

First workflow

  1. Add the server to an MCP client using the install method above.

  2. Call create_brain with a supported YouTube channel URL, such as https://www.youtube.com/@OpenAI.

  3. Immediately call get_brain_status once with wait_until_terminal=true and keep the same turn active until it returns.

  4. Use search_brain to retrieve timestamped caption matches.

The initial ingestion scans the complete channel listing so it can select up to 50 videos by view count. Caption availability and YouTube rate limits determine how much can be indexed.

Configuration

Local data and privacy

By default, Channel Brains stores its SQLite database and lock file in the platform's user-data directory:

  • Windows: %LOCALAPPDATA%\channel-brains-mcp

  • macOS: ~/Library/Application Support/channel-brains-mcp

  • Linux: ~/.local/share/channel-brains-mcp

Set CHANNEL_BRAINS_HOME before launching the MCP server to use another location:

CHANNEL_BRAINS_HOME=/path/to/channel-brains-data uvx --from "git+https://github.com/Pu11en/channel-brains@v0.1.4" channel-brains-mcp

Captions and search indexes stay on the local machine. The only network requests are public YouTube requests made by yt-dlp and caption URL retrieval during ingestion.

YouTube rate limits

Channel Brains uses paced requests and bounded retries. If YouTube keeps limiting requests, the brain pauses with its progress saved. Wait a while, then call create_brain again with the same channel URL to resume.

Manual MCP client configuration

The automated install above covers Codex, Claude Code, ZCode, and Hermes. For any other MCP client, the generic server definition is:

{
  "command": "uvx",
  "args": [
    "--from",
    "git+https://github.com/Pu11en/channel-brains@v0.1.4",
    "channel-brains-mcp"
  ]
}

Use a startup/connect timeout of at least 60 seconds. The first uvx run downloads the pinned package and can be slower than later starts.

Per-client configuration snippets (Hermes YAML, Codex TOML, ZCode/OpenCode JSON) are in docs/clients.md.

For AI agents

The automated agent installation flow is fully specified in AGENT_INSTALL.md. If you are an agent that has been asked to install this repository, read that file completely and execute the matching client procedure. Do not start YouTube ingestion during installation; wait for an explicit channel URL and indexing request.

Development and verification

uv sync --extra dev --locked
uv run ruff check .
uv run pytest
uv build

The default test suite is fully offline. The manual live release gate exercises a real channel through caption ingestion and timestamped local search:

uv run pytest -m live tests/test_live_youtube.py

Override the default channel when needed:

CHANNEL_BRAINS_LIVE_CHANNEL=https://www.youtube.com/@OpenAI uv run pytest -m live tests/test_live_youtube.py

YouTube requests are paced. HTTP 429 responses receive bounded retries and then pause the brain without losing completed work. Call create_brain again later with the same channel URL to resume.

Limitations

  • Public YouTube captions can be unavailable, expired, restricted, or rate-limited. No local client can guarantee that an external YouTube request succeeds 100% of the time; Channel Brains guarantees bounded behavior, resumable progress, and explicit status when YouTube refuses a request.

  • Search is lexical SQLite FTS5 search, not semantic search or an answer-generation system.

  • The server indexes captions only. It does not download videos, reuse video footage, or create a knowledge graph.

  • A channel can contain many videos. The first local ingestion may take time.

Brand assets

The brand mark is the red video badge with a white top-view brain ("Top Brain"). Files live in logos/final/:

Concept history from the design rounds is in the git history.

License

MIT. © Drew Pullen. See LICENSE.

Available Tools

6 tools
create_brainB

Explicitly queue or resume local channel caption indexing.

Inputs: channel_url (required), language, max_videos. Returns: queued, status, message, brain_id, +4 more

delete_brainA

Permanently delete one brain only with confirm=true.

Inputs: brain_id (required), confirm. Returns: deleted, message, brain_id, deleted_chunk_count, +1 more

get_brain_statusA

Read progress for one brain, or all local brains. Never contacts YouTube.

Inputs: brain_id. Returns: count, brains

get_video_transcriptA

Page through indexed caption chunks for one video. Never contacts YouTube.

Inputs: brain_id (required), video_id (required), limit, offset. Returns: limit, total, chunks, offset, +5 more

list_brain_videosB

Page through selected videos and outcomes. Never contacts YouTube.

Inputs: brain_id (required), limit, offset. Returns: limit, total, offset, videos, +2 more

search_brainB

Return ranked, timestamped caption evidence. Never generates an answer or contacts YouTube.

Inputs: query (required), limit, brain_id. Returns: query, results, brain_id, brain_statuses, +2 more

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool targets a distinct operation: creation, status, listing, search, transcript retrieval, and deletion. No two tools overlap in purpose, so agents can easily distinguish them.

    Naming Consistency5/5

    All tools follow a consistent verb_noun snake_case pattern (e.g., create_brain, delete_brain, get_video_transcript). The repeated verb 'get' is used clearly for different resource types.

    Tool Count5/5

    Six tools form a compact, focused set for a YouTube caption indexing server. Each tool serves a clear role without bloat or redundancy.

    Completeness5/5

    The server covers the full lifecycle: create (index) a brain, check status, list videos, search captions, retrieve transcripts, and delete. No essential operation appears missing for its stated purpose.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers