claude-reader

A live reading pane for Claude Code sessions. It shows only the text Claude writes to you, stripped of tool calls, bash output and diffs, with a sidebar to jump back to any earlier message.

Works in any terminal, next to any Claude Code session, with no plugins and no changes to Claude Code itself. (Independent project, not affiliated with Anthropic.)

┌────────────────────────────────┬───────────────────────────────────────────────┐
│ claude-reader — Improve message readability  [5e86140a]                        │
├────────────────────────────────┼───────────────────────────────────────────────┤
│ Can we make this a TUI?        │  Claude · 07:25:25                            │
│ you · 07:25:01                 │  ─────────────────────────────────────────    │
│ You're right that a web page…  │                                               │
│ claude · 07:25:25              │  You're right that a web page adds friction,  │
│ So I run it in a second pane?  │  and the terminal-native version is just as   │
│ you · 07:27:37                 │  feasible. But one framing correction first…  │
│ Almost. One nuance to get…     │                                               │
│ claude · 07:27:50              │  - tmux split                                 │
│ Let's build it.                │  - your terminal's native split               │
│ you · 07:30:15                 │  - or a second window                         │
│ ▶ Setting up in the project…   │                                               │
│ claude · 07:30:38              │  So the deliverable is one command…           │
└────────────────────────────────┴───────────────────────────────────────────────┘

Why

In a Claude Code session, the prose that matters to you as the human (findings, questions, summaries) is buried in a stream of tool calls and command output. Scrolling back to re-read an earlier message means paging through hundreds of lines of noise, and the terminal never renders Claude's markdown.

claude-reader fixes that by tailing the session transcript that Claude Code already writes to disk. It keeps the assistant text blocks and your own prompts, drops everything else, and renders the result as markdown in a two-pane TUI that updates in real time.

Install

With pipx or uv:

pipx install claude-reader
# or
uv tool install claude-reader

Or run without installing:

uvx claude-reader

(From source: pipx install git+https://github.com/vikasgrac/claude-reader.)

Requires Python 3.10+ on Linux/macOS (or WSL). The only dependency is Textual. Tests: pip install -e . pytest pytest-asyncio && python -m pytest.

Use

Open a second pane next to your Claude Code session. Any pane works: a tmux split, your terminal's native split (Ghostty, Kitty, iTerm2, Warp, Windows Terminal hosting WSL or an ssh session), or a second window. Then, from the same directory the Claude session was started in, run:

claude-reader

It finds the newest session for that project and follows it. If several sessions in the project were active in the last hour, a picker asks which one you want.

Keys:

Key	Action
↑/↓ or k/j	move through messages (reading pane follows)
f	jump to the latest message and resume auto-following
s	show/hide the sidebar for full-width reading
u	show/hide your own prompts
c	copy the current message to the clipboard
p	session picker: switch to another session in this project
q	quit

Auto-follow pauses whenever you select an older message, so new messages keep appending to the sidebar without pulling you away from what you're reading.

Options:

claude-reader path/to/transcript.jsonl   # open a specific transcript
claude-reader --no-pick                  # skip the startup session picker
claude-reader --mouse                    # enable in-app mouse (see below)
claude-reader --max-messages 200         # keep only the newest N messages (default 500)

Unknown options and extra arguments are usage errors (exit 2); a transcript that cannot be opened is reported on stderr (exit 1). Only the newest --max-messages messages are kept in memory and shown in the sidebar, so a very long transcript opens fast and stays cheap.

Two-minute tour

1. Start Claude Code in a project as usual: cd ~/work/myapp && claude.
2. Split your terminal (tmux: prefix %; or a second tab/window) and in the new pane run:

   cd ~/work/myapp && claude-reader
   

   The sidebar fills with the conversation so far, newest at the bottom, and the reading pane shows the latest message rendered as markdown.
3. Ask Claude something long — a code review, a plan. While it works, the tool calls scroll by in the Claude pane; in the reader nothing moves until Claude writes prose, and then that message appears, formatted, with headings, lists and code blocks intact.
4. Press k a few times to walk back through earlier messages. Auto-follow pauses while you read; press f to jump back to the newest and resume following.
5. Press u to hide your own prompts and see only Claude's side, s to hide the sidebar and read full width, c to copy the current message to your clipboard for a note or a commit message.
6. Ran several sessions in this project today? p opens the session picker; the live ones are marked. Pick one and the reader switches to it.
7. q quits. Claude Code never noticed any of this — the reader only ever read a file.

How it works

Claude Code logs every session to ~/.claude/projects/<project>/<session-id>.jsonl, one JSON object per line, normally appended to. Each assistant turn is a list of typed blocks: text, tool_use, thinking. claude-reader keeps only the text blocks (plus your prompts, with injected system context stripped), which is exactly the separation you want. It remembers its byte offset into the file and re-reads whatever was appended, twice a second, so it stays live with no hooks into Claude Code or the terminal. If the file is truncated or replaced (it tracks size and inode) it starts over and reloads; records it does not understand — nulls, new record types, a half-written last line, stray bytes — are skipped, never fatal.

Because it only reads a file, it is not tied to any terminal app, works over ssh, and does not break when Claude Code's UI changes.

Notes for tmux and ssh users

Clipboard. c copies through tmux load-buffer -w when running inside tmux, which forwards to the outer terminal's clipboard via OSC 52 (a temporary named tmux buffer is used and deleted right after, so the text does not linger in tmux's paste history). This works over ssh if your terminal supports OSC 52 (most do). Outside tmux it uses Textual's clipboard support. Copying is the only time text leaves the app: it reads transcripts locally and makes no network calls.

Selecting text in the Claude pane. Two things in a neighbouring TUI can break drag-selection in the Claude Code pane, and claude-reader avoids both by default:

• Terminal mouse tracking. Textual normally enables any-motion mouse reporting. claude-reader runs with the mouse off unless you pass --mouse. Under tmux the scroll wheel still moves through messages, because tmux converts wheel events to arrow keys for full-screen apps that do not track the mouse.
• Periodic output. A ticking clock in the header would repaint the pane every second, and output landing under an in-progress drag kills the selection. claude-reader draws nothing unless a new message arrives.

Status

0.2.0 — used daily since day one, hardened after an outside review: malformed or half-written transcript lines are skipped, truncated/replaced files reload, memory is bounded (--max-messages), and there is a test suite. Things it does not do yet: search, a compressed trace of tool activity between messages, a claude-watch wrapper that opens Claude and the reader in one tmux command. Issues and PRs welcome.

License

MIT