SmallClawLM

Zero-token AI agent powered by Google NotebookLM.

One agent, one notebook, one specialty. Each SmallClawLM agent owns exactly one NotebookLM notebook — the notebook's sources define its domain expertise. All reasoning runs through Gemini inside NotebookLM. No external LLM API keys needed.

Architecture: Hybrid Fast/Slow Path

SmallClawLM uses an intent router to classify user input:

• Fast Path (Pipeline): Direct tool calls for known intents — podcast, report, quiz, research, etc. No LLM reasoning needed. Saves rate limits for actual work.
• Slow Path (Agent): Full CodeAgent with NLMModel brain for conversational, analytical, or multi-step tasks that need reasoning.

User Input
    │
    ▼
┌──────────┐     ┌──────────────────┐
│  Router   │────►│ Fast Path: Pipeline │──► Direct API Call ──► Result
│ (intent)  │     └──────────────────┘
│           │────►┌──────────────────┐
│           │     │ Slow Path: Agent   │──► NLMModel.chat ──► CodeAgent ──► Result
└──────────┘     └──────────────────┘

Agent Specialties

Specialty	Tools	Use Case
research	deep_research, ask_notebook, add_source, list_sources	Web research and Q&A
podcast	ask_notebook, add_source, generate_podcast	Audio overviews
quiz	ask_notebook, add_source, generate_quiz	Study aids
report	ask_notebook, add_source, generate_report	Structured reports
mindmap	ask_notebook, add_source, generate_mind_map	Visual summaries
all	All tools	General purpose

Install

pip install smallclawlm
# Or from source:
git clone https://github.com/aissac/smallclawLM.git
cd smallclawLM
pip install -e .

Authenticate

smallclaw login
# Opens browser for Google authentication

Usage

CLI — Auto-routed (recommended)

# Auto-routes to fast path (direct podcast generation)
smallclaw run "generate a podcast"

# Auto-routes to fast path (deep research)
smallclaw run "research fusion energy breakthroughs"

# Auto-routes to slow path (conversational reasoning)
smallclaw run "Why is cold fusion so difficult to achieve?"

# Force slow path (agent reasoning)
smallclaw run "explain quantum computing" --force-slow

# Force fast path
smallclaw run "generate report" --force-fast

# Verbose routing info
smallclaw run "research fusion energy" -v

CLI — Shortcut Commands (fast path only)

smallclaw research "mRNA vaccine technology"
smallclaw podcast --notebook-id abc123
smallclaw report --notebook-id abc123
smallclaw quiz --notebook-id abc123
smallclaw mindmap --notebook-id abc123
smallclaw list-sources --notebook-id abc123

CLI — Pipeline

# Chain operations without an agent
smallclaw pipe --add-source "https://arxiv.org/abs/2401.12345" --generate podcast
smallclaw pipe --research "climate change" --generate report
smallclaw pipe --ask "What are the key findings?" --notebook-id abc123

CLI — Interactive Agent

# Start a slow-path session
smallclaw agent --notebook-id abc123

Python API

from smallclawlm import NLMAgent, create_agent, route

# Auto-route: fast or slow path
result = route("generate a podcast")
# RouteResult(path=Path.FAST, intent="generate_podcast", confidence=0.9, params={})

# Fast path: direct Pipeline call
from smallclawlm.extensions.pipeline import Pipeline, ArtifactType
pipe = Pipeline(notebook_id="abc123")
pipe.research("fusion energy")
result = pipe.execute()

# Slow path: full agent
agent = create_agent("research", notebook_id="abc123")
result = agent.run("Why is cold fusion difficult to achieve?")

Design Decisions

• Hybrid Architecture: Fast path for known intents (saves rate limits), slow path for reasoning
• One Agent = One Notebook: No multi-brain routing. NotebookLM is a reading assistant
• Stateless Chat: conversation_id=None. smolagents manages ReAct history
• Pipeline Pattern: Declarative chaining without LLM calls
• Error → string conversion: Tools return errors as strings so CodeAgent can self-correct

Error Handling

Error	Recovery
ChatError (rate limit)	Exponential backoff
NetworkError (timeout)	Retry with backoff
ValueError (auth expired)	Auto refresh + retry

License

Apache-2.0