Gab n' Go

Treat the LLM context window like a filesystem. Concepts (constraints, goals, preferences, observations, references) are files you push into context and pull back out.

The context is just JSON — llcat conversation files. Gab n' Go is the curated store; push injects concepts as system messages into those files, pull scans them and imports what it finds. No FUSE, no magic — just cp-style sync on JSON.

Concept Schema

Each concept is a JSON object with these fields:

Field	Limits
type	constraint, goal, preference, observation, reference
description	Under 50 tokens
short	Under 250 characters
medium	Under 1000 characters
long	Under 2500 characters

Example:

{
  "type": "constraint",
  "description": "programming languages",
  "short": "The programming language that we use is C",
  "medium": "We use C17 standard and gcc-13 with a compile target of AMD64. Our build system is cmake"
}

Project Structure

File	Purpose
index.html	Web interface — drag-and-drop concept management
server.py	HTTP server serving the UI + API to concepts.json
concept.py	CLI for listing and filtering concepts (Typer)
concept_mcp.py	MCP server exposing concepts as a tool
config.py	Shared llama.cpp model configuration
concepts.json	Sample concept data
schema	Concept field constraints
instruct.txt	Prompt to get an LLM to output concept JSON
tester.html	HTML drag-and-drop prototype for the concept window
pyproject.toml	Project metadata and build configuration
requirements.txt	Python dependencies

Dependencies

pip install -r requirements.txt

• typer – CLI framework
• mcp – Model Context Protocol SDK
• rich – terminal formatting
• chromadb – vector database for concept storage

Usage

CLI — browse, push, pull

# List all concepts
python concept.py main concepts.json

# Filter by type
python concept.py main concepts.json --type constraint

# Print as an LLM prompt (inject into context window)
python concept.py main concepts.json --type preference --format llm

# Push concepts into an llcat conversation JSON (as system message)
python concept.py push conversation.json

# Push with a specific detail level
python concept.py push conversation.json --scope medium

# Pull concepts from an llcat conversation JSON into the store
python concept.py pull conversation.json

Web Interface — browse, drag-and-drop, import/export

python server.py

Open http://localhost:8080 in a browser. The web UI lets you:

• Paste concept JSON or load from file
• Browse concepts with type filters
• Drag-and-drop to reorder
• Delete individual concepts
• Export the full concept list back as JSON

MCP Server — query concepts from any MCP host

python concept_mcp.py

Runs over stdin/stdout. Clients call get_concept with optional type and description to pull concepts into context.

LLM Integration

Use instruct.txt to prompt an LLM to summarize conversation topics as concept JSON. Paste the output into the web interface or save it to concepts.json.

Roadmap

• Concept CRUD CLI + MCP tool
• ChromaDB vector storage
• Basic web interface for browsing and drag-and-drop concept management
• Push/pull sync — push concepts into llcat conversation JSON, pull them back
• Web interface push/pull — push/pull from the browser

Configuration

See config.py for model path, context size, GPU layers, and generation params. Default: models/qwen2.5-3b-instruct-q2_k.gguf.