beyin 0.3.4

Build local, queryable packs from videos, articles, podcasts, and files for MCP and local LLM use.

beyin

base engine for your information nodes

also means “brain” in Turkish.

Build local, queryable packs from videos, articles, podcasts, and local files.
Query them through MCP with your AI agent, or explore them directly with a local model.

✨ Features

• 🔗 MCP compatible: works with Claude Code, Codex, Cursor, Windsurf, Zed and more
• 📦 Local-first pipeline: processing, embedding, and storage all happen on your machine
• 🎬 Rich source support: YouTube videos and playlists, podcasts, PDFs, articles, local files
• 🌍 50+ languages: multilingual embedding model out of the box
• 🤖 Ollama support: run fully offline with a local model
• ⚡ Plug and play: one command to connect via MCP, then manage everything by just talking to your agent
• 🎯 Multi-query expansion: generates query variants automatically for better retrieval

⚙️ How it works

The recommended way to use beyin is through MCP with the AI agent you already use.

1. Install beyin and connect it to your agent once
2. Tell your agent to create a pack, add your sources, and build it
3. Ask questions naturally and let your agent handle retrieval

It becomes part of your agent’s natural workflow, bringing in the right context exactly when you need it while keeping you focused

See Example Usage with MCP for real prompts and workflows.

You can also query packs directly with a local model, no external API or agent needed. See Query with a Local Model.

📂 Supported Sources

Type	Examples
Web articles	Public URLs
YouTube	Videos and playlists
Podcasts	RSS feed URLs
Local documents	.pdf, .docx, .pptx, .epub, .xlsx, .csv
Local text	.txt, .md, .rst, .html
Local audio	.mp3, .m4a, .wav
Local video	.mp4, .mov, .mkv, .webm

beyin is built for local processing on your own machine. Use it with content you are allowed to process, preferably public, permitted sources or material you own or have rights to use. Avoid copied, paywalled, private, restricted, or illegally shared content.

📦 Installation

We recommend uv because it keeps the CLI install isolated and gives you a plain beyin command on your PATH.

If you do not have uv yet, install it first:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Package manager alternatives also work, for example brew install uv on macOS or winget install --id=astral-sh.uv -e on Windows. See the official uv installation guide for more options: docs.astral.sh/uv/getting-started/installation.

The recommended way to install beyin is with uv:

uv tool install beyin

If you already manage a dedicated Python environment and explicitly want beyin inside it, pip install beyin still works, but it is a fallback path rather than the recommended default.

ffmpeg is required for video and audio sources. Skip if you only use articles and local files:

# macOS
brew install ffmpeg

# Linux
sudo apt install ffmpeg

# Windows
winget install ffmpeg

No Homebrew on macOS or winget not working? Download directly from ffmpeg.org/download.html.

Windows note

If you are on Windows and plan to use beyin through Codex MCP, prefer WSL over native Windows. See Windows notes for the recommended setup and details.

🖥️ Using via CLI

You can use beyin directly in your terminal without MCP:

beyin

If you want to verify runtime dependencies before building anything:

beyin check-deps

What happens on first run:

• beyin starts a guided setup flow
• after setup, if you do not have any packs yet, beyin shows a start screen where you can create a new pack or import an existing one
• if setup is already complete and no packs are available, that same start screen is shown again
• if you already have packs, you can manage and build them from the CLI as usual

Useful examples:

beyin
beyin help
beyin list
beyin build my-pack
beyin settings
beyin check-deps

If you do not want a persistent install, you can run the same commands with uvx beyin ... instead.

🔌 Connect to Your Agent

You only need to do this once.

beyin runs as a stdio MCP server. After registration, your MCP client launches and manages the server for you, so you do not need to keep a separate terminal open. If the tool does not appear immediately, restart the client.

Claude Code

claude mcp add beyin -- beyin mcp-server

To make it available across all your projects:

claude mcp add --scope user beyin -- beyin mcp-server

Codex

codex mcp add beyin -- beyin mcp-server

Cursor

Open or create ~/.cursor/mcp.json and add:

{
  "mcpServers": {
    "beyin": {
      "command": "beyin",
      "args": ["mcp-server"]
    }
  }
}

Or go to Command Palette → "View: Open MCP Settings".

Windsurf

Open or create ~/.codeium/windsurf/mcp_config.json and add:

{
  "mcpServers": {
    "beyin": {
      "command": "beyin",
      "args": ["mcp-server"]
    }
  }
}

Or go to Command Palette → "MCP: Add Server".

Zed

In ~/.config/zed/settings.json:

{
  "context_servers": {
    "beyin": {
      "source": "custom",
      "command": "beyin",
      "args": ["mcp-server"]
    }
  }
}

Any other MCP-compatible agent

Recommended command:

beyin mcp-server

Use this as the MCP server command in any client that supports the MCP protocol. Your client will launch and manage it for you, so you do not need to run it manually in a separate terminal.

💬 Example Usage with MCP

Once beyin is connected through MCP, you can talk to your agent naturally. You do not need to memorize commands or even say "beyin" every time. Just ask for what you want.

Some prompts that mention local files or folders may require your AI agent to have read access to those locations first.

What you want	What to say
Build a new pack	create a pack called yt-research, add this YouTube playlist: https://youtube.com/playlist?list=..., and build it
Add a source	I have a PDF about growth strategy in my Downloads folder, add it to my mobile-marketing pack and rebuild
Add more sources	add these to my product ideas pack and rebuild: https://example.com/article-1, https://example.com/article-2, https://example.com/article-3
Ask a question	any useful info about onboarding screens in my mobile marketing pack?
Control the response	ask yt-research pack about building an audience from scratch, include sources and timestamps
Check your packs	list my packs and show me their status
Ask about a pack	whats the status of mobile marketing pack? and what about its sources?
Remove a source	remove sources 2 and 3 from mobile marketing pack
Remove a pack	remove that pack about tech podcast

🛠️ MCP Tools Reference

These are the tools beyin exposes to your agent. Your agent uses them automatically; you do not need to call them yourself.

Tool	What it does
packs	List all installed packs
status	Show details and readiness for a pack
retrieve	Return relevant results for one or more queries
build	Build or update a pack. Pass sources to build only selected sources by index or range. Automatically purges chunks of removed sources.
add	Add a pack from a path, URL, or YAML
add_sources	Add new sources to a pack. Rebuilds automatically for single sources; playlists/feeds are expanded for review first.
remove_sources	Remove sources by index, range, or text match. Removed chunks stay in the vector store until you rebuild.
remove	Remove an installed pack (moves to trash)
registry	Browse the beyin community registry by topic, tag, or keyword

📋 All Commands

Pack lifecycle

Command	What it does
beyin create	Create a new pack interactively
beyin add <path-or-url>	Import an existing pack from a file or URL
beyin build <pack>	Build or rebuild a pack
beyin update <pack>	Fetch new content and rebuild incrementally
beyin remove <pack>	Remove a pack
beyin list	List all installed packs
beyin status <pack>	Show pack details and readiness

Sources

Command	What it does
beyin add-source <pack> <url>	Add a new source to an installed pack
beyin remove-source <pack> <selector>	Remove a source by index, range, or text match

Query

Command	What it does
beyin query <pack> "question"	Ask a question directly (requires Ollama)

Server & config

Command	What it does
beyin mcp-server	Start the MCP server
beyin settings	View and configure settings
beyin check-deps	Verify runtime dependencies
beyin about	Version and info
beyin help	List all commands

Common Examples

Build

Command	What it does
beyin build <pack>	Build all current pack sources
beyin build <pack> --source 2	Build one selected source by index
beyin build <pack> --source 1-3	Build a selected source range
beyin build <pack> --source "keyword"	Build selected sources by visible text match
beyin build <pack> --model tiny --language tr	Override Whisper model and language for this build
beyin build <pack> --reset	Clear runtime artifacts and rebuild from scratch

Add Sources

Command	What it does
beyin add-source <pack> <url>	Add one source
beyin add-source <pack> <url-1> <url-2>	Add multiple sources in one command
beyin add-source <pack> ./notes.md	Add a supported local file path as a source
beyin add-source <pack> <url> --build	Add source(s) and build immediately

Remove Sources

Command	What it does
beyin remove-source <pack> 2	Remove source by index
beyin remove-source <pack> 1 3 5	Remove multiple sources by index
beyin remove-source <pack> 1-3	Remove a range of sources
beyin remove-source <pack> "keyword"	Remove a source by title or URL text match
beyin remove-source <pack> 2 --build	Remove and rebuild immediately to clean up vector store

🤖 Query with a Local Model

You can query your packs with a local model using Ollama, without sending anything to an external API. Everything stays on your machine.

If you use beyin through an MCP-connected agent (Claude Code, Codex, etc.), you do not need Ollama. Your agent is the LLM. beyin just retrieves results for it.

Setup:

1. Download and install Ollama from ollama.com
2. Pull a model:

# Example models
ollama pull llama3.2     # 2 GB, fast, good for most queries
ollama pull qwen2.5:7b   # 4.7 GB, stronger reasoning

3. Start Ollama in separate terminal window:

ollama serve

4. Query with beyin in another terminal:

beyin query my-pack "What does this source say about X?"

To change the model, run beyin settings.

🔧 Troubleshooting

• Pack is not queryable yet
• MCP is connected but retrieval is not working
• Audio and video builds are slow
• Video or audio builds fail
• Pack name with spaces is not recognized
• Which install path should I use?
• Windows notes

Pack is not queryable yet

beyin status my-pack
beyin build my-pack

A partially-ready pack is still queryable — sources that built successfully are available. Rebuilding recovers any failed sources.

MCP is connected but retrieval is not working

• Make sure the pack was built: beyin status my-pack
• Restart your agent after adding beyin for the first time
• Verify the server is registered: claude mcp list
• Make sure the same beyin installation is used by both CLI and the MCP server

Audio and video builds are slow

beyin uses Whisper to transcribe audio and video sources. Model size controls the trade-off between speed, memory use, and transcription quality. English-only .en variants through medium.en are useful when you know the audio is only English.

Model	Type	Parameters	Download size	Speed	Accuracy	Best for
tiny	multilingual	39M	~75 MB	fastest	lowest	Quick tests, clean audio, mixed-language detection
tiny.en	English-only	39M	~75 MB	fastest	low	Fastest English-only transcripts
base	multilingual	74M	~145 MB	fast	low	Simple podcasts, lightweight multilingual audio
base.en	English-only	74M	~145 MB	fast	low+	English podcasts and interviews
small	multilingual	244M	~483 MB	moderate	good	Most use cases, multilingual content
small.en	English-only	244M	~483 MB	moderate	good+	Strong default for English-only speech
medium	multilingual	769M	~1.5 GB	slow	better	Harder English, multilingual, accented, or noisy audio
medium.en	English-only	769M	~1.5 GB	slow	better+	Higher English accuracy without multilingual support
large	multilingual	1.55B	~3 GB	slowest	best	Maximum accuracy, difficult audio

The default model is small. To use a faster or English-only model, change it in settings:

beyin settings

Or pass it per build:

beyin build my-pack --model small

small is a good default for most content. If your audio is strictly English, small.en is a good faster/simpler option. Use medium, medium.en, or large for harder audio.

Video or audio builds fail

• Check that ffmpeg is installed: ffmpeg -version
• Check that yt-dlp is installed and current: yt-dlp --version
• Make sure the source URL is still reachable

Pack name with spaces is not recognized

Pack IDs use kebab-case, not spaces. Use my-pack instead of my pack. The display name can be anything, but the ID used in commands must be kebab-case.

Which install path should I use?

Prefer one of these two patterns and stay consistent:

• Installed workflow: uv tool install beyin, then run beyin ...
• No-install workflow: run uvx beyin ...

Why uv is the main path:

• uv tool install installs beyin as a standalone CLI app instead of dropping it into whatever Python environment happens to be active
• you get a plain beyin command on your PATH, so day-to-day usage stays simple
• the install is isolated, which avoids environment drift and version clashes from unrelated Python projects
• it is easier to keep CLI usage and agent/MCP usage aligned when beyin is installed as one managed tool

Do not mix them casually across CLI and MCP setup, or you may end up checking one environment and running another. pip install beyin is available if you explicitly want beyin inside a managed Python environment, but it is not the main path.

Windows notes

The main beyin ... commands stay the same on Windows after installation. The examples above use Unix-like shell syntax and config paths in a few places, so if a path does not match your system, use the equivalent settings location for your MCP client on Windows. For media builds, install ffmpeg first, then rerun beyin check-deps.

If you are using Codex through MCP on Windows, prefer WSL over native Windows. In testing, native Windows CLI installs could work, but Codex MCP calls were noticeably less reliable there, with long cold starts or repeated stuck first calls when using both beyin mcp-server and uvx beyin mcp-server. WSL was much more consistent.

Recommended Windows setup for Codex:

1. Install WSL and open your Linux distro terminal.
2. Install uv inside WSL.
3. Install beyin inside WSL with uv tool install beyin.
4. Add the MCP server from the same WSL environment with codex mcp add beyin -- beyin mcp-server.

🧑‍💻 Development

git clone https://github.com/buralog/beyin.git
cd beyin
uv sync

Run commands from the repo:

uv run beyin help

MCP config for a local repo install:

claude mcp add beyin -- uv run beyin mcp-server --cwd /absolute/path/to/beyin

Or manually in your agent's config file:

{
  "mcpServers": {
    "beyin": {
      "command": "uv",
      "args": ["run", "beyin", "mcp-server"],
      "cwd": "/absolute/path/to/beyin"
    }
  }
}

Run tests:

uv run pytest tests/test_cli.py tests/test_mcp_server.py

🔍 Behind the Scenes

1. beyin fetches or loads your source content
2. It extracts text or generates transcripts (for audio/video)
3. It chunks the content into indexed segments
4. It embeds those chunks into a local vector store
5. At query time, it retrieves the best-matching chunks using multi-query expansion

beyin uses a multilingual embedding model by default, so it works well across 50+ languages.

Privacy note: Steps 1–4 are entirely local. At step 5, only the retrieved chunks reach your LLM. For full privacy, use beyin with Ollama so nothing leaves your machine.

🤝 Contributing

Issues and pull requests are welcome at github.com/buralog/beyin.

See CONTRIBUTING.md for pack submissions, pack policy, and code contribution guidelines.

⚖️ Legal

beyin does not host, publish, or redistribute third-party content. Any retrieval, transcription, indexing, or embedding of source material happens locally on the end user's own machine.

Users are responsible for ensuring that their use of beyin complies with applicable laws, copyright rules, and the terms of service of the source platforms.

📄 License

MIT