Phi-4-mini • 3.8B params • 8 scaffolding layers • $0.00 inference cost
Autonomous AI agents that run entirely local — no cloud, no API keys, no credit card.
How wide is the gap between your model's raw capability and what your project actually does? A Tier 1 model producing Tier 3 results wins. Every time.
Achieved entirely through scaffolding — not model capability. The model provides raw text. The scaffolding provides structure, validation, memory, recovery, and amplification.
Each layer compensates for a specific weakness of the 3.8B model. Together they form a robust agent framework that rivals cloud-based systems — at zero cost.
Detects tool execution errors via keyword matching + result inspection. Injects healing guidance. Auto-retries up to 3 cycles with alternative approaches.
Screenshot → analyze → click/type/scroll → repeat. 8-cycle autonomous web navigation. No pre-parsed HTML needed.
Breaks complex tasks into JSON plans with dependency graphs. Executes in topological order with final synthesis step.
Sequential tool execution with step tracking, cross-referencing, and circuit breakers (max 10 steps, 60s timeout).
Extracts factual claims and verifies them against source context. Annotates output: made / verified / unverified / blocked. Falls back to "I don't know."
SQLite-backed long-term memory with semantic search. Survives restarts. Importance scoring + category filtering. 10,000-entry capacity.
3 perspective agents (Optimist, Pessimist, Analyst) debate in parallel on the same 4B model. Consensus synthesis with confidence score.
On failure, model introspects: "Why did this fail?" Generates improved strategy and retries. Up to 3 refinement cycles with comparison tables.
Beyond individual agents — Ultraclaw orchestrates swarms of parallel AI agents and interfaces with physical robot hardware through the OpenClaw skill registry.
Multi-agent swarm: parallel sub-agents execute tasks independently, then synthesize results
Speed + direction control for wheeled rovers
Text-to-speech output via robot voice modules
Camera capture + vision analysis pipeline
Facial expression display on robot screens
LIDAR scanner for obstacle detection (10m range)
Each extension is registered in the openclaw_skills.rs registry — plug-and-play for edge AI deployments.
From home automation to team collaboration — here's what Ultraclaw handles autonomously.
Control smart devices through the SmartHome skill — lights, thermostats, sensors. All local, no cloud dependency.
Vision-based navigation: screenshot → analyze → click/type/scroll. Works on any rendered page without pre-parsed HTML.
Native connectors for Discord, Matrix, Telegram — each with isolated session context per room/channel.
Git operations, file system analysis, conflict resolution. The GitResolver skill handles merge conflicts autonomously.
Image generation via DALL-E/Imagen, TTS via Whisper, and screenshot-based vision analysis.
Runs on Raspberry Pi, $500 laptops, air-gapped systems. Zero cloud requirement. Zero API keys.
Parallel sub-agents (Analyst, Coder, Researcher, Reviewer) execute tasks independently, then synthesize results.
100% local execution — no data ever leaves your machine. Full offline capability with zero telemetry.
Every problem below is addressed directly by Ultraclaw's engineering — not just the model.
Frontier models are rate-limited and expensive — inaccessible to most builders globally.
Phi-4-mini runs entirely local via Ollama. No API keys. No rate limits. No monthly bills.
API budgets that cost hundreds per month exclude billions of builders from the AI revolution.
CPU-only. 6GB RAM. A student in a developing country can run the same agent as a funded startup.
Cloud endpoints don't reach offline devices. Air-gapped, rural, and edge use cases are abandoned.
No internet needed after initial model download. Works in basements, farms, and remote locations.
Cloud-based agents send all your data to external servers — chats, files, system info.
100% local. No telemetry. Landlock kernel sandbox constrains all network + filesystem access.
Small models hallucinate freely — generating plausible-sounding but completely fabricated claims.
Every factual claim verified against source. Unverifiable claims blocked. Falls back to "I don't know."
When small models produce bad output, there's no second chance — the task simply fails.
Errors detected automatically → healing guidance injected → retried with alternative approach. Up to 3 cycles.
☁ TYPICAL CLOUD AGENT
🦀 ULTRACLAW LOCAL
| Feature | Ultraclaw (Local Phi-4-mini) | Typical Cloud Agent |
|---|---|---|
| Setup cost | $0 | $50–200/month |
| Hardware required | $500 laptop (CPU only) | GPU server |
| Offline capable | ✅ Yes | ❌ No |
| Privacy | 100% local | Data leaves machine |
| Streaming output | ✅ Yes | ✅ Yes |
| Tool calling | ✅ Yes | ✅ Yes |
| Multi-turn conversations | ✅ Yes | ✅ Yes |
| Self-healing | ✅ Auto-retry (3 cycles) | ❌ No |
| Persistent memory | ✅ SQLite + semantic search | ❌ Stateless |
| Multi-agent debate | ✅ 3-perspective consensus | ❌ No |
One command setup. No API keys. No cloud account. No credit card. Runs on any $500 laptop with 6GB available RAM.
Ultraclaw is written in Rust and compiled natively. Install the Rust toolchain first:
Restart your terminal after installation. Verify with rustc --version and cargo --version.
Ollama runs Phi-4-mini locally on CPU — no GPU required.
Windows users: Download the installer directly from ollama.com/download. Run the .exe and follow the installation wizard. macOS users can also use brew install ollama.
Download the 3.8B parameter model. Fits in 4GB of RAM — runs on CPU only.
After download, verify with ollama list — you should see phi4-mini:latest in the list.
Clone the Ultraclaw source code from GitHub:
Compile the Rust binary with full optimizations. First build may take 5-10 minutes depending on your CPU.
The optimized binary (~15MB) will be at target/release/ultraclaw (Linux/Mac) or target\release\ultraclaw.exe (Windows).
Release profile uses: LTO (link-time optimization), single codegen unit, panic=abort, strip — all tuned for minimal binary size and maximum performance.
Three ways to launch Ultraclaw depending on your needs:
🎮 Demo Mode — Hackathon Showcase (recommended first run):
Runs through all 8 scaffolding layers in a single narrative flow. Shows per-layer metrics, memory stats, self-healing demonstrations, debate synthesis, and a final cost comparison dashboard.
⌨ Interactive CLI Mode:
Full interactive chat interface. Type commands, ask questions, and watch the agent execute tools in real-time with streaming output.
📺 TUI Mode — Terminal Dashboard:
Keyboard-driven Ratatui interface with live LLM heartbeat monitoring, swarm routing displays, and real-time inference statistics.
Create a .env file in the Ultraclaw directory. Copy from the example:
Default configuration (works out of the box for demo mode):
Optional additions for real deployments:
Build with platform-specific connectors via Cargo features for real-world deployment:
Each connector runs with isolated session context — a conversation in one room/channel can never leak into another. Least-privilege scopes per connector.
TECHNICAL_WRITEUP.md.One command to rule them all. No API keys. No cloud. No cost.