On July 14, Anthropic released Claude Code v2.1.208 with an accessibility feature and a list of fixes for issues reported by the community. The main addition is the screen reader opt-in mode: users of screen readers can enable plain-text rendering with claude --ax-screen-reader, an environment variable, or a flag in the config file. The version also fixes critical crashes (background sessions that wouldn’t restart after a binary update, stream-json blocking on Windows-style empty lines) and automatically restores fast mode when switching back to a model that supports it.

Why it matters to you. If you use Claude Code daily, this release removes real friction: users running background agents no longer lose response text when delivery fails, those writing long prompts in stream-json won’t see truncated output anymore, and anyone with visual impairment can now use the tool with their screen reader. More importantly: the fix list is long and specific, a sign that Anthropic is closing gaps reported from the field instead of just pushing new features. As we covered on July 11, after the July 5 security incident Claude Code had changed its architecture to set auto mode as default on three platforms; this release is the follow-up: stability and accessibility first.

If you want to try it. If you already use Claude Code, the update is automatic. To enable screen reader mode: claude --ax-screen-reader from the terminal, or add "axScreenReader": true to your settings file.

Details

Release v2.1.208 is the first after switching to auto mode as default (July 11) and comes with twenty-four entries in the release notes, almost all fixes. Three main areas:

Accessibility: the screen reader mode

The new feature is plain-text rendering for screen reader users, activable in three ways: --ax-screen-reader flag from command line, CLAUDE_AX_SCREEN_READER=1 environment variable, or "axScreenReader": true key in the JSON config file. There are no public screenshots or demos, but the expected behavior is rendering without complex visual formatting (no boxes, interactive progress bars, graphical menus) that common screen readers struggle to interpret.

It’s an explicitly opt-in mode, not automatic detection: those who want it must enable it. This avoids Claude Code trying to guess whether a screen reader is in use (non-trivial across all systems) and leaves control with the user.

Crashes and state loss

The most substantial fixes concern crash scenarios reported in production:

• Text typed to a background agent that couldn’t deliver it was permanently lost. Now the text is saved and delivered on session restart.
• The background daemon failed to attach after an automatic binary update (the parent process was still running the old version). Now handled.
• Long prompts in stream-json mode (piping from SDK or external scripts) produced truncated output or missing responses due to unflushed buffers. Fixed.
• Supervised and background sessions crashed when a server closed an HTTP/2 connection with GOAWAY while requests were in flight. Now handled with graceful degradation.

The full list includes eleven other fixes for edge cases (files modified after reading, invalid regex in Grep, massive markdown tables blocking render), but the four above were the most reported on GitHub Issues.

Implications for those building with Claude Code

This release adds no new agentic functionality, and that’s intentional: after the integrated browser on July 12 and auto mode on July 11, Anthropic is focusing on staying stable under load. The fixes section is nearly triple the features, and many fixed bugs had been open for weeks. This is the typical pattern of a tool moving from demo phase to production: fewer announcements, more robustness.

The most interesting detail is the screen reader mode as explicit opt-in rather than automatic detection. Conservative choice, but honest: better an explicit option that works than heuristics that fail half the time. Users with visual disabilities will find it documented, those without it see no changes.

Where to look

The complete release notes on GitHub list all fixes with issue numbers. If you use Claude Code in an automated workflow (scripts, CI, or unsupervised agents), it’s worth reading the section on stream-json and background daemon: the fixes for state loss and output truncation directly address those use cases.