Skip to content

TRUST.md

Latest commit

 

History

History
39 lines (30 loc) · 3.01 KB

TRUST.md

File metadata and controls

39 lines (30 loc) · 3.01 KB
Raw

The code-stick Trust Model

code-stick is built for environments where security, privacy, and host integrity are non-negotiable. This document outlines the five pillars of trust that ensure your host machine remains clean, your code stays private, and the binaries you run are exactly what you expect.

1. Zero Residue (Environment Isolation)

The "Zero Install" promise is enforced by strict environment isolation. When you launch code-stick from a USB:

  • Redirected Configs: All application configurations are redirected to <USB>/config.
  • Isolated Caches: Temporary files and npm caches are pinned to <USB>/cache.
  • Local State: Runtime logs and internal state are stored in <USB>/state.
  • Result: No files are written to your AppData, Documents, or ~/.config folders. Your host remains as clean as it was before you plugged in the stick.

2. 100% Offline (Airgapped by Design)

The tool is designed to work in restricted environments (banks, defense, hospitals).

  • No Telemetry: There are zero background "pings," analytics, or heartbeat calls.
  • Pre-staged Dependencies: Every runtime dependency needed for the AI agent to work is pre-staged onto the USB during the initial install.
  • Local Inference: All model weights live on the USB and run on the host's RAM/GPU without ever reaching out to the internet.

3. Privacy-Preserving Bug Reports

We believe you shouldn't have to leak your identity to report a bug.

  • Automatic Redaction: Our crash reporter automatically scrubs your home directory path, hostname, USB drive letters, and common API key patterns (sk-, npm_, ghp_) before the report ever touches your disk.
  • Manual Upload Only: Reports are saved to your local temp directory for your review. code-stick will never auto-upload a report. You are in total control of what you share.

4. Binary Integrity & Pinned Hashes

Every binary we ship is a verifiable "unit of trust."

  • SHA-256 Verification: The installer pins every version of Ollama and opencode with a cryptographic SHA-256 hash.
  • Tamper Protection: If a download is corrupted or tampered with in transit, the installer will immediately detect the mismatch and abort the installation.
  • Verifiable Pins: You can cross-check our pins directly against the upstream GitHub releases for both projects.

5. Transparent & Verifiable

Security should not be a "black box."

  • Open Source: The entire code-stick logic is under 5,000 lines of TypeScript, making it easy to audit in an afternoon.
  • Plain Text Launchers: The scripts that start the agent (.bat, .sh, .command) are plain text. You can open them in any editor to see exactly how environment variables are being set and which processes are being spawned.
  • Self-Audit Guide: Our SECURITY.md file provides a "cheat sheet" of where to look in the source code to verify every security claim we make.

For a deep dive into the technical implementation and threat model, see SECURITY.md.