terminal · harness · project

agent-tty (Coder)

A terminal an agent drives, with every session recorded and exportable as text snapshots, PNG screenshots, asciicasts or WebM, so a human can verify what it did.

complementevidence: strongby Codergithub.com/coder/agent-tty ↗

agent-tty (Coder)

  • Maker: Coder
  • URL: https://github.com/coder/agent-tty
  • Status (2026-10-10): 30 stars, Apache-2.0, TypeScript, Node 24 or later. Created 19 March 2026, last push 24 September 2026; npm 0.5.0 (23 June 2026), about 185 downloads last month.

What it is

“Give your AI agent a real terminal it can drive, and get back reviewable snapshots, screenshots, and recordings of everything it did.” A PTY driver whose distinguishing feature is proof: every session is recorded and can be exported as a text snapshot, PNG, asciicast or WebM.

The problem it’s solving

You can’t take an agent’s word that it tested the TUI it just changed. agent-tty makes the agent attach evidence a person can check at a glance, and gives a live dashboard to watch sessions.

Its path / bet

Playwright’s loop for terminals (create, run, wait, snapshot, screenshot, destroy), with reviewability as the product. Screenshots and video render through Playwright’s Chromium.

How it works

JSON output on every command; wait --text and --screen-stable-ms; send-keys with chords; record export --format webm; a “doctor” check; dogfood transcripts of Codex and Claude using it.

Strengths

  • Clear thesis (evidence, not claims) and good agent-facing CLI design.
  • Backed by a company that sells agent infrastructure.

Weaknesses / limits

  • Heavy dependencies (Node 24, Chromium for pictures).
  • Same semantic blindness as every driver.
  • Little adoption so far.

Relation to fictty

Complement and inspire. fictty’s planned history (record every input, export, replay) is the same idea from inside the runtime: an agent session on a fictty screen should produce a replayable record without a separate recorder.

Could fictty adopt it instead of building?

No for the screen. For demo and PR evidence, using agent-tty on a fictty screen works today, but fictty’s own history will be more exact (states, not pixels).

What fictty should take from it

  • Make “attach proof to the PR” a first-class export of fictty history: a .cast, a text transcript of frames, and a picture.
  • A doctor command that checks the terminal (kitty graphics, mouse, size) and says what will degrade.

Sources