star-terminal

A personal AI assistant that runs on your machine, powered by Nova models on AWS Bedrock.

star-terminal is an always-on desktop assistant organized around projects and skills.

A project is an isolated workspace — "manage my Linux servers", "personal finance", "home automation" — with its own chat history, scheduled prompts, and sandbox permissions. Each project defines exactly what the AI is allowed to do: which hosts it can reach, which paths it can read or write, whether it can run shell commands, and which environment variables it can see.

Skills are a global library of small, sandboxed programs that the AI creates, validates, and executes on your behalf. You describe what you need in natural language; star-terminal interviews you, generates the code, security-scans it, and runs it inside a permission-controlled Deno Worker. Skills are independent of any single project — the same skill can be enabled across many projects, and each project can enable many skills. Before a skill can run in a project, the user must explicitly consent to it; consent is scoped to the (skill, project) pair. The effective permissions a skill gets at runtime are the intersection of what the skill declares and what the project allows.

What it does

  • Conversational AI chat — multi-turn chat backed by AWS Bedrock, with full tool-use support
  • Custom skills — describe a task (e.g. "check my server uptime", "summarise this RSS feed") and star-terminal builds, scans, and runs a skill for it
  • Built-in system skills — web search (Brave / Tavily / DuckDuckGo), shell commands, HTTP fetch, and push notifications (ntfy.sh / Slack)
  • Scheduled prompts — set a cron schedule for any project and star-terminal runs your prompt automatically in the background
  • Project isolation — each project has its own chat history, enabled skills, sandbox permissions, and schedule
  • Security by default — every generated skill is AI-scanned before activation; execution is sandboxed with per-project permission controls (network, filesystem, shell, env)
  • Skill versioning — full version history with one-click revert; code is regenerated from the text snapshot, not stored verbatim
  • Persistent daemon — installs as an OS service (launchd / systemd) so scheduled tasks and background skills run even when the TUI is closed
  • Encrypted secrets — API keys and credentials stored with AES-GCM encryption in SQLite

Tech stack

Deno + TypeScript, SQLite for persistence, AES-GCM encryption for secrets, Deno Workers for sandboxed skill execution. Ships as a single compiled binary — no runtime dependencies.

Architecture

The daemon owns the database, runs the AI agent loop, and executes skills. The TUI connects over a local TCP socket and provides an interactive terminal interface with keyboard-driven navigation.

Built With

  • aes-gcm-encryption-for-secrets
  • bedrock
  • cli
  • deno
  • nova
  • sqlite
  • sqlite-for-persistence
  • typescript
Share this project:

Updates

Submission history