Architecture Review: Sled

Sled claims to be Run your coding agent from your phone, with voice. Let’s look under the hood.

🛠️ The Tech Stack

Sled differentiates itself from standard “chat with code” apps by focusing on secure, local connectivity rather than cloud hosting. It is essentially a remote control for your local terminal.

• Connectivity Layer (The “Deep” Part): Instead of uploading your repo to a Sled server, it utilizes Tailscale to create a secure, peer-to-peer mesh network. This allows the phone to talk directly to your local development machine without exposing ports to the open internet.
• The Agent Host: It runs a local server on your machine that interfaces with CLI-based agents. It explicitly supports Claude Code, OpenAI Codex, and Gemini CLI. This means the “intelligence” is brought by your existing tools, while Sled handles the I/O.
• Voice Pipeline: The mobile app handles Voice-to-Text (Input) and Text-to-Speech (Output). It captures your voice command, tunnels it to your local machine, executes it against the running agent (e.g., “Fix the linting errors in main.py”), and streams the text response back as audio.
• Privacy Posture: High. Since the code execution happens locally and traffic is tunneled via Tailscale, your codebase technically never leaves your controlled environment (except for the API calls your agent makes to the LLM provider).

💰 Pricing Model

Open Source / Free Sled is positioned as a fully open-source tool.

• Self-Hosted: You run the host on your machine.
• BYOK (Bring Your Own Key): You pay for the underlying API usage (Anthropic/OpenAI) via the agents you run (like Claude Code).
• Infrastructure: Relies on the free tier of Tailscale for personal use.

⚖️ Architect’s Verdict

Sled is a Sophisticated Wrapper that solves a specific infrastructure problem: AFK (Away From Keyboard) latency.

While it doesn’t train its own models (making it a wrapper around Claude Code/LLMs), it acts as a critical “Bridge” technology. Most coding agents today require you to stare at a terminal to approve changes or iterate. Sled decouples the developer from the chair.

Developer Use Case:

1. The “Dog Walking” Deploy: You trigger a long-running migration or test suite. You leave the desk. Sled reads you the error logs via TTS, and you verbally command the agent to apply a fix while walking outside.
2. Idea Capture: You verbally scaffold a new feature while commuting, having the agent create the files on your local machine so they are ready when you arrive.

It is Production Ready for individual developers who are comfortable with CLI tools and Tailscale, but teams should be wary of voice-coding critical production hotfixes without visual confirmation.