SmallestClaw

SmallestClaw is a voice-first autonomous agent — you call a phone number, talk to it, and it does things on your computer. No screen, no keyboard, no app. Just your voice.

Comment

🦞 SmallestClaw

📞 Call your computer. Literally.

SmallestClaw is a voice-first autonomous agent — you call a phone number, talk to it, and it does things on your computer. No screen, no keyboard, no app. Just your voice.

"Hey SmallestClaw, what's my git status?" "Read me the last error in the server logs." "Deploy the staging branch."

🤔 Why?

Every AI assistant today is text-first. You type, it types back. Voice is bolted on as an afterthought — a glorified dictation layer over a chat window.

SmallestClaw flips this. The phone call IS the interface. There is no chat window. There is no screen to look at. You pick up the phone and talk to your computer like it's a person sitting at the keyboard.

This isn't voice-as-input. This is voice-as-product.

⚡ How it works

📞 You call in
   ↓
🎙️ Smallest.ai handles real-time voice (STT → LLM → TTS in <800ms)
   ↓
🧠 GLM-5 (via Featherless.ai) decides what to do
   ↓
🦞 OpenClaw executes it autonomously on your machine
   ↓
🔊 Results spoken back to you naturally

🎙️ Powered by Smallest.ai — The Voice Engine

Smallest.ai is the backbone of SmallestClaw. This isn't a project that happens to use voice — it's a project that couldn't exist without Smallest.ai's voice infrastructure.

Here's what Smallest.ai's Atoms platform makes possible:

  • ⚡ Pulse STT — Speech-to-text in 64ms. When you speak, the agent is already processing before you finish your sentence. This is what makes the conversation feel real, not like talking to a speakerphone with a 3-second delay.
  • 🗣️ Sub-800ms turn latency — From the moment you stop talking to the moment the agent starts responding. Faster than most human conversations. This is the difference between a voice assistant you tolerate and one you actually want to use.
  • 📞 Native telephony — Real phone numbers, real phone calls. Not a browser demo, not a "press the microphone button" widget. You call a number and talk. Smallest.ai handles the entire telephony stack — PSTN routing, audio codecs, echo cancellation — so we can focus on the agent logic.
  • 🔊 Natural TTS — The voice that speaks back to you doesn't sound like a robot reading a script. Smallest.ai's synthesis models produce natural, expressive speech that makes long interactions comfortable.
  • 🏗️ Atoms SDK — The Python framework that ties it all together. Session management, event handling, streaming responses, tool coordination — all the plumbing that would take months to build from scratch is handled by the SDK. We wrote ~200 lines of agent code. Atoms handles the other 200,000.

Without Smallest.ai, SmallestClaw would be a chatbot with a microphone bolted on. With Smallest.ai, it's a voice-native product where the phone call IS the experience.

🧠 Powered by Featherless.ai — The Open Brain

SmallestClaw doesn't use a closed-source LLM behind a walled garden. It runs on Featherless.ai — open model inference that gives us the power of frontier models without the lock-in.

  • 🔓 Open models, zero lock-in — We're running GLM-5 by ZAI, an open-weight model served through Featherless's OpenAI-compatible API. Switch models with one line of config. No vendor lock-in, no "we changed our pricing" surprises.
  • 🔌 Drop-in compatible — Featherless exposes a standard OpenAI-compatible API (/v1/chat/completions). The Smallest.ai Atoms SDK's OpenAIClient connects to it natively — just change the base_url. No adapter code, no translation layer, no hacks.
  • 🛠️ Tool calling support — GLM-5 via Featherless supports function calling, which is critical for SmallestClaw. When you say "check my git status", the LLM doesn't just generate text — it emits a structured tool call that triggers real execution on your machine. Without reliable tool calling, the entire agent loop breaks down.
  • 💰 Cost-effective inference — Open model inference through Featherless is dramatically cheaper than closed APIs, making it feasible to run a voice agent that might handle dozens of turns per call without burning through budget.

Featherless makes SmallestClaw's brain portable, affordable, and open — while still being smart enough to decide when to run a command vs. ask a clarifying question.

🛠️ What it can do

Say this... SmallestClaw does this
"Check my git status" 🖥️ Runs git status, reads you the summary
"What's in my package.json?" 📄 Reads the file, tells you the key details
"Search for the latest Next.js release" 🌐 Searches the web, summarizes results
"Set up a new Python project with FastAPI" 🤖 Creates directories, writes files, installs deps — autonomously
"Find and fix the broken import in src/" 🔧 Scans files, identifies the issue, applies the fix

It doesn't just answer questions. It does things.

🏗️ Architecture

┌──────────────┐     ┌──────────────────┐     ┌──────────────┐
│   📞 Phone   │────▶│  🎙️ Smallest.ai  │────▶│  🧠 GLM-5    │
│   (You)      │◀────│  Voice Pipeline   │◀────│  Featherless │
└──────────────┘     └────────┬─────────┘     └──────┬───────┘
                              │                       │
                              │    tool calls         │
                              ▼                       │
                     ┌──────────────────┐             │
                     │  🦞 OpenClaw     │◀────────────┘
                     │  Local Agent     │
                     │  (your machine)  │
                     └──────────────────┘
  • app.py — Entry point. Wires the Atoms voice session to OpenClaw.
  • agent.py — The voice agent. Defines tools (run_command, read_file, search_web, do_task) that dispatch to OpenClaw.

🚀 Quick start

# Install
uv sync

# Configure (fill in your API keys)
cp .env.example .env

# Make sure OpenClaw is running locally, then:
uv run python app.py

# In another terminal — talk to it:
smallestai agent chat

📖 Full deployment guide (including phone number setup): DEPLOY.md

🎯 What makes this different

  • 🎙️ Voice-first, not voice-added — The phone is the product. No fallback UI.
  • 🦞 Actual autonomy — Doesn't just talk about doing things. OpenClaw does them.
  • ⚡ Real-time — Smallest.ai's Pulse transcription in 64ms + sub-800ms turn latency. Feels like talking to a person.
  • 🔓 Local execution — OpenClaw runs on YOUR machine. Your files never leave your network.
  • 🧠 Open model — GLM-5 via Featherless, no closed-API dependency.

🏛️ Built with

  • Smallest.ai Atoms — Voice AI platform with real-time STT/TTS and native telephony
  • Featherless.ai — Open model inference with OpenAI-compatible API
  • OpenClaw — Open-source autonomous AI agent
  • GLM-5 by ZAI — The open LLM powering the brain

Built With

Share this project:

Updates