⚛️ AgentNova R02.6

A minimal, hackable agentic framework engineered to run entirely locally with Ollama or BitNet.

Inspired by the architecture of OpenClaw, rebuilt from scratch for local-first operation.

Written by VTSTech · GitHub

---

📚 Documentation

Document	Description
Architecture.md	Technical documentation for developers (directory structure, core design, orchestrator modes)
CHANGELOG.md	Version history and release notes (includes LocalClaw history)
TESTS.md	Benchmark results, model recommendations, and testing guide

---

Installation

# Install from GitHub using pip:
pip install git+https://github.com/VTSTech/AgentNova.git

Backward Compatibility

The package was previously named localclaw. For backward compatibility:

# Old package name still works (shows deprecation warning)
pip install localclaw

# Old CLI command still works
localclaw run "What is the capital of Japan?"  # Redirects to agentnova

# Old imports still work (with deprecation warning)
import localclaw  # Re-exports from agentnova

We recommend updating to the new package name:

# Old
import localclaw
from localclaw import Agent

# New
import agentnova
from agentnova import Agent

From Source

git clone https://github.com/VTSTech/AgentNova.git
cd AgentNova
pip install -e .

No Installation Required

AgentNova uses only Python stdlib — no dependencies! You can also just copy the agentnova directory into your project:

cp -r agentnova /path/to/your/project/

---

Quick Start

1. Test Model Tool Support (Recommended First Step)

# Test all models for native tool support
agentnova models --tool_support

# Results saved to tested_models.json for future reference

2. Single prompt

# Simple Q&A
agentnova run "What is the capital of Japan?"

# With streaming output
agentnova run "Tell me a joke." --stream

# Specify a model
agentnova run "Explain quantum computing" -m llama3.2:3b

3. Interactive chat

# Start interactive session
agentnova chat -m qwen2.5-coder:0.5b

# With tools enabled
agentnova chat -m llama3.1:8b --tools calculator,shell,read_file,write_file

# With skills loaded
agentnova chat -m llama3.2:3b --skills skill-creator --tools write_file,shell

# Fast mode (reduced context for speed)
agentnova chat -m qwen2.5-coder:0.5b --fast --verbose

4. Using BitNet backend

agentnova chat --backend bitnet --force-react
agentnova run "Calculate 17 * 23" --backend bitnet --tools calculator

---

Key Features

• Zero dependencies — uses Python stdlib only
• Ollama + BitNet backends — switch with --backend flag
• Three-tier tool support — native, ReAct, or none (auto-detected per model)
• Agent Skills — follows Agent Skills specification
• Small model optimized — pure reasoning mode for sub-500M models
• Built-in security — path validation, command blocklist, SSRF protection

---

Tool Support Levels

AgentNova automatically detects each model's tool support level:

Level	Description	When to Use
native	Ollama API tool-calling	Models trained for function calling
react	Text-based ReAct prompting	Models that accept tools but need format guidance
none	No tool support	Models that reject tools; use pure reasoning

Testing Tool Support

# Test all models
agentnova models --tool_support

# Example output:
  Model                                      Family       Context    Tool Support
  ──────────────────────────────────────────────────────────────────────────────
  gemma3:270m                                gemma3       32K        ○ none
  granite4:350m                              granite      32K        ✓ native
  qwen2.5-coder:0.5b-instruct-q4_k_m         qwen2        32K        ReAct
  functiongemma:270m                         gemma3       32K        ✓ native

Performance by Tool Support

R02.6 Quick Diagnostic results (5 questions, ~30-120s/model):

Model	Score	Time	Tool Support
functiongemma:270m	100%	19.6s	native
granite4:350m	100%	49.4s	native
qwen2.5:0.5b	100%	66.3s	native
qwen2.5-coder:0.5b	100%	116.5s	react
qwen3:0.6b	100%	122.8s	react
gemma3:270m	80%	14.3s	none
dolphin3.0-qwen2.5:0.5b	80%	26.6s	none
qwen:0.5b	20%	27.0s	none

Key improvements in R02.6:

• 5 models achieve 100% - All tool-calling models now perfect!
• functiongemma:270m fastest at 19.6s (native tools + 270M params)
• Multi-step expression extraction - Handles 8 times 7 minus 5, word problems, time calculations
• ReAct JSON parsing fixed - Clean extraction even with trailing text
• Verbose response fallback - Uses numeric result when model gives long explanation
• Tool-calling outperforms pure reasoning - ALL native/react models score 100% vs max 80% for no-tool models

---

CLI Commands

Command	Description
run "prompt"	Run single prompt and exit
chat	Interactive multi-turn conversation
models	List available Ollama models with tool support info
tools	List built-in tools
skills	List available Agent Skills
test [example]	Run example/test scripts (--list to see all)
modelfile [model]	Show model's Modelfile system prompt

Key Flags

Flag	Description
-m, --model	Model name (default: qwen2.5-coder:0.5b)
--tools	Comma-separated tool list
--skills	Comma-separated skill list
--backend	ollama or bitnet
--stream	Stream output token-by-token
--fast	Preset: reduced context for speed
-v, --verbose	Show tool calls and timing
--acp	Enable ACP (Agent Control Panel) integration
--use-mf-sys	Use Modelfile system prompt instead of AgentNova default
--force-react	Force ReAct mode for all models
--debug	Show debug info (parsed tool calls, fuzzy matching)
--num-ctx	Context window size for test commands
--num-predict	Max tokens to predict for test commands

Models Command

# List models with family, context size, and tool support
agentnova models

# Test each model for native tool support (recommended)
agentnova models --tool_support

Output shows:

• Model - Model name
• Family - Model family from Ollama API
• Context - Context window size
• Tool Support - ✓ native, ReAct, ○ none, or untested

⚛️ AgentNova R02.6 Models
  Model                                      Family       Context    Tool Support
  ──────────────────────────────────────────────────────────────────────────────
  gemma3:270m                                gemma3       32K        ○ none
  granite4:350m                              granite      32K        ✓ native
  qwen2.5-coder:0.5b-instruct-q4_k_m         qwen2        32K        ReAct
  functiongemma:270m                         gemma3       32K        untested

  1 model(s) untested. Use --tool_support to detect native support.

Test Command Examples

# List all available tests
agentnova test --list

# Quick diagnostic - 5 questions, ~30s/model (NEW in R02.3)
agentnova test 15 --model granite3.1-moe:1b
agentnova test 15 --model all --debug

# Run GSM8K benchmark (50 math questions)
agentnova test 14 --acp --timeout 6400

# Run with debug output
agentnova test 02 --debug --verbose

---

Built-in Tools

Tool	Description
calculator	Evaluate math expressions
python_repl	Execute Python code
shell	Run shell commands
read_file	Read file contents
write_file	Write content to file
list_directory	List directory contents
http_get	HTTP GET request
save_note / get_note	Save and retrieve notes

---

Configuration

Variable	Description	Default
OLLAMA_BASE_URL	Ollama server URL	http://localhost:11434
BITNET_BASE_URL	BitNet server URL	http://localhost:8765
ACP_BASE_URL	ACP (Agent Control Panel) server URL	http://localhost:8766
AGENTNOVA_BACKEND	Backend: ollama or bitnet	ollama
AGENTNOVA_MODEL	Default model	qwen2.5-coder:0.5b-instruct-q4_k_m
AGENTNOVA_SECURITY_MODE	Security mode: strict, permissive, disabled	permissive

---

Setup Ollama

# Make sure Ollama is running:
ollama serve

# Pull a model:
ollama pull qwen2.5-coder:0.5b-instruct-q4_k_m

# Test tool support:
agentnova models --tool_support

---

About

**⚛️ AgentNova ** is written and maintained by VTSTech.

• 🌐 Website: https://www.vts-tech.org
• 📦 GitHub: https://github.com/VTSTech/AgentNova
• 💻 More projects: https://github.com/VTSTech

---

For more details, see:

• Architecture.md — Technical architecture and design decisions
• CHANGELOG.md — Version history and release notes
• TESTS.md — Benchmark results and model recommendations