Claude Code has always been a power tool for developers who want an AI coding agent that operates directly in the terminal. What it has not been, until now, is accessible to developers who rely on screen readers such as VoiceOver on macOS or NVDA on Windows. Anthropic has shipped a dedicated screen reader mode that fundamentally changes how Claude Code renders its output — and the engineering behind it is more considered than a simple accessibility checkbox.
What Screen Reader Mode Actually Does
The core problem with terminal AI interfaces and screen readers is one of rendering: Claude Code's default interface uses box-drawing characters, in-place redraws, progress animations, and spinners that are either meaningless noise or actively disorienting when read aloud. Screen reader mode strips all of that away.
With the mode active, Claude Code writes flat, linear text. No box-drawing characters. No color-only cues. No content redraws. Progress spinners become static text. Tables in Claude's replies — which would normally render as a box-character grid — are reformatted as Header: value sentences. Output accumulates in the terminal's scrollback buffer, which means a developer can use their screen reader's review commands to re-read any part of a previous turn without hunting through visual chrome.
Every section of the transcript gets a labeled prefix: you: for your messages, claude: for replies, tool: for file edits or commands, tool error: for failures, and Permission Required: for any prompt that needs approval. Those labels are also searchable, giving screen reader users a reliable way to jump between sections of a long session.
Three Ways to Turn It On
Anthropic has built in three activation methods with a clear precedence order, which matters for developers managing Claude Code across different environments — including remote SSH sessions.
For a single session, the --ax-screen-reader flag is the fastest path: claude --ax-screen-reader. For all sessions started from a given shell, setting CLAUDE_AX_SCREEN_READER=1 as an environment variable — and adding it to a shell profile — covers every new terminal. For a machine-wide setting that includes the VS Code integrated terminal, "axScreenReader": true in the user settings file is the right approach.
The mode confirms its own activation on startup with a labeled line: [Screen Reader Mode: on via flag], [Screen Reader Mode: on via env], or [Screen Reader Mode: on via settings], so there is no ambiguity about whether it is running. On SSH connections, the environment variable or settings method is recommended since the flag applies only to the local invocation.
Screen reader mode requires Claude Code v2.1.181 or later. The table-reformatting feature — arguably the most practically useful change for longer sessions — requires v2.1.198.
Permission Prompts and Menu Navigation
One of the sharper details in the implementation is how permission prompts are handled. In standard Claude Code, these are navigated with arrow keys. In screen reader mode, they become numbered lists. Each option is read as a numbered line, followed by an input prompt that specifies the valid range. A developer types the number and presses Enter rather than attempting to navigate a visual menu.
Yes/no prompts accept typed y or n input instead of a two-option menu. Dismissible menus end with an or Escape to cancel instruction. The terminal bell rings when Claude finishes a reply, when a permission prompt appears, and when a tool that has been running for more than five seconds completes — meaning a developer using a screen reader does not have to continuously monitor the transcript to know when attention is needed.
For developers working across long sessions, Claude Code also emits OSC 133 shell-integration markers at turn boundaries. In terminals that support them — iTerm2 on macOS, VS Code's integrated terminal, Windows Terminal — this enables jump-to-previous-prompt navigation without reading through the entire transcript. This is a meaningful quality-of-life feature for any session with significant back-and-forth, not just accessibility use cases.
Additional Accessibility Settings
Screen reader mode is not the only accessibility lever Anthropic has exposed. CLAUDE_CODE_ACCESSIBILITY=1 keeps the native terminal cursor visible for screen magnifier users — relevant for developers using macOS Zoom or similar tools who do not need a screen reader but do need cursor tracking. The prefersReducedMotion setting disables spinners and animations without otherwise changing the interface. The theme setting includes dark-daltonized and light-daltonized colorblind-friendly themes.
These settings work alongside screen reader mode rather than replacing it, which means a developer with overlapping needs — a screen magnifier and reduced motion preferences, for instance — can combine them without conflict.
Known Gaps
Anthropic has been transparent about what is not yet handled. Screen reader mode does not activate automatically when a screen reader is detected — it requires explicit opt-in. Attaching to a background session via claude attach enters the terminal's alternate screen, which has no native scrollback; the recommendation is to press Left Arrow on an empty prompt to exit. Session costs are announced only in the exit summary, not per turn.
These are real limitations, and Anthropic has pointed developers toward the Claude Code issue tracker to report issues with specific assistive technologies. The fact that the known limitations section exists and is specific is a reasonable signal that the team expects to iterate.
For a tool that has not been without its reliability issues, getting the fundamentals of accessibility right is a meaningful step — and one that opens Claude Code to a segment of the developer community that has largely been excluded from the agentic coding wave so far.
