Recoleta
GitHub
Source note

Show HN: AI integrated in any terminal that's invisible until you need it

Authors
emosenkis
Source
Hacker News
Published
Collection
Software Intelligence
Terminal AICode IntelligenceDeveloper ToolsMCPHuman AI Interaction
Open originalMarkdown source

Summary

Terminai wraps a normal terminal with an on-demand AI overlay for Codex, Claude Code, or another CLI agent. Its main claim is safer terminal assistance through read access and approval-gated writes, while leaving model auth and routing to the user's existing AI CLI.

Problem

  • Terminal AI agents need shell context, but direct write access can be risky when commands affect files, processes, or remote systems.
  • Many users already have preferred AI CLIs, model credentials, and provider settings, so a terminal helper that forces a new provider path adds friction.
  • The product matters for developer workflows because terminal work often mixes inspection, command generation, debugging, and file changes in one session.

Approach

  • Terminai runs as a wrapper around the user's shell or another command, started with terminai or terminai -- <command> [arg1 arg2...].
  • Pressing Ctrl+Space opens an overlay terminal that runs the selected AI CLI.
  • The wrapped shell exposes context through a Terminai MCP server, giving the agent read access and controlled write access.
  • Write actions are approval-gated, so queued shell input from the agent requires user approval before it reaches the wrapped terminal.
  • Presets are bundled for Codex and Claude Code, while custom CLIs can be used if they can consume the MCP URL and context prompt.

Results

  • No benchmark, user study, latency metric, accuracy metric, or task-completion comparison is provided in the excerpt.
  • The product claims support for 3 agent paths: Codex, Claude Code, and a user-provided CLI agent.
  • It claims support for 2 operating systems: Linux and macOS.
  • It claims Terminai makes 0 outgoing network connections and does not collect user data; the configured AI CLI handles authentication and model selection.
  • Installation paths include Homebrew, source build with cargo install --path src, and GitHub releases.
  • The author labels the software alpha-quality and reports using it as a daily driver, with native scrolling and input handling named as difficult implementation areas.

Link