Untether engine compatibility

Untether supports four coding agents. The basics work with all of them. The interactive controls depend on what each agent's CLI exposes, so Claude Code gets the most.

  • Claude Code

    Complex refactors, architecture and long context. Plans, questions, diff preview and background tasks are Claude Code only.

    npm i -g @anthropic-ai/claude-code
  • Codex CLI

    Fast edits, shell commands and quick fixes, on your ChatGPT subscription. Codex runs without stopping to ask, so you pick full auto or a read-only safe mode before each run.

    npm i -g @openai/codex
  • OpenCode

    Over 75 providers through Models.dev, plus local models.

    npm i -g opencode-ai@1
  • Pi

    Sign in with several providers, with a conversational style.

    npm i -g @mariozechner/pi-coding-agent

Other engines

  • Antigravity CLIPlanned for v0.36.1

    Google's successor to Gemini CLI. Its own antigravity engine is planned for Untether v0.36.1.

  • Gemini CLIDeprecated

    Still included and still runs, but no longer supported, tested or fixed, and it may be removed later. Google retired Gemini CLI for individual accounts on 18 June 2026.

  • AmpDeprecated

    Still included and still runs, but no longer supported. Amp refuses clients it considers out of date, so a working setup can stop without notice.

Works with every engine

These work the same whichever of the four supported engines you use.

Live progress
Tool calls, file edits and elapsed time stream into the chat as your agent works.
Resume and continue
Reply to continue a session, or send /continue to pick up one from your terminal. Pi needs provider = "openai-codex" for subscription sign-in.
Switch models
/model overrides the model for an engine, per chat, and the footer shows which model ran.
Voice notes
Voice notes are transcribed and sent to your agent as text.
Reply with context
Reply to a message, or quote part of it, and your agent sees that text.
Verbose progress
/verbose shows each tool call's details as it runs.
Plain errors
Outages, rate limits and billing errors arrive as plain messages with resume steps.
Projects and branches
Switch repos, run a branch in its own worktree, or give each forum topic its own project.
Files both ways
Upload files into the repo, browse folders, and get results back as documents.
Scheduled tasks
Cron schedules, webhooks and one-off /at runs can start any engine.
Export and stats
/export saves a transcript, and /stats counts runs and time per engine.
Settings menu
/config switches the engine and model with buttons. Some engines add more options.

Where engines differ

Everything not listed above depends on the engine.

Yes Partly (see the note) No

Where engines differ
Claude CodeCodex CLIOpenCodePi
Approvals and plans
Approve tool calls with buttonsYesNoNoNo
Plan mode and plan outlinesYesNoNoNo
Questions with option buttonsYesNoNoNo
Diff preview before editsYesNoNoNo
Auto-approve safe toolsYesNoNoNo
Approval policy set before a runYesYes, see note 1NoNo
Sessions
Background tasks and live sessionsYesNoNoNo
Steer messages into a running sessionYesNoNoNo
Loop modeYesNoNoNo
Context use in the status lineYesNoNoNo
Context compaction shownYesNoNoYes
Models and costs
Reasoning and effort levelsYesYesNoNo
Permission mode shown in the footerYesPartly, see note 3NoNo
Cost tracking and budgets in dollarsYesPartly, see note 2YesPartly, see note 2
Subscription quota in /usageYesNoNoNo
Sign in again from Telegram (/auth)NoYesNoNo
  1. Codex runs in full auto (its own sandbox setting) or safe mode (a read-only sandbox, so edits are blocked). You choose in /config before a run; Codex can't stop to ask mid-run.
  2. Codex and Pi report token counts, not dollars, so dollar budgets don't apply to them. OpenCode reports a cost on paid providers and token counts on its free tier.
  3. Codex shows safe mode in the footer. Full auto runs show no mode.

Gemini CLI and Amp are left out because they're deprecated. The full table, including them, is in the Untether README.