claude-replay

Community tool — not affiliated with or endorsed by Anthropic.

AI coding sessions are great for development, but hard to share. Screen recordings are bulky and transcripts are hard to navigate.

claude-replay turns Claude Code, Cursor, Codex CLI, Gemini CLI, OpenCode, and Kimi Code session logs into interactive, shareable HTML replays. The generated replay is a single self-contained HTML file with no external dependencies — you can email it, host it anywhere, or embed it in documentation. Use --serve --watch to monitor agent sessions live as they run.

Try it online · Live demo

Claude Code, Cursor, Codex CLI, Gemini CLI, OpenCode, and Kimi Code store conversation transcripts on disk. claude-replay auto-detects the format and converts them into visual replays suitable for blog posts, demos, and documentation.

Features

• Self-contained HTML output (no dependencies)
• Interactive playback with speed control
• Collapse/expand tool calls and thinking blocks (Claude's internal reasoning traces)
• Bookmarks / chapters
• Secret redaction before export
• Multiple color themes
• Terminal-style bottom-to-top scroll
• File activity sidebar — see which files were touched and navigate to their tool calls
• Embeddable via iframe
• Live watch mode — monitor agent sessions in real time (--serve --watch)
• Web-based editor UI for visual session editing and preview

Use cases

claude-replay is useful for:

• Blog posts — show AI-assisted development sessions interactively
• Documentation — embed AI debugging sessions or code walkthroughs
• Demos — share reproducible sessions without video
• Bug reports — attach a replay instead of long logs
• Teaching — step through AI reasoning and tool usage
• Live monitoring — watch agent sessions in real time on remote machines or containers

Installation

npm install -g claude-replay

Or run directly with npx (zero install):

npx claude-replay

Docker

docker run --rm --init -p 7331:7331 \
  -v ~/.claude/projects:/root/.claude/projects:ro \
  ghcr.io/es617/claude-replay

Open http://localhost:7331 for the web editor. Session directories are mounted read-only.

For CLI usage:

docker run --rm --init \
  -v ~/.claude/projects:/root/.claude/projects:ro \
  -v $(pwd):/output \
  ghcr.io/es617/claude-replay \
  /root/.claude/projects/my-project/session.jsonl -o /output/replay.html

Quick start

# Launch the web editor (default)
claude-replay

# Generate a replay by session ID (auto-finds the file)
claude-replay abc123def456 -o replay.html

# Or pass the full path
claude-replay ~/.claude/projects/-Users-me-myproject/session-id.jsonl -o replay.html

# Chain multiple sessions into one replay
claude-replay session1-id session2-id -o combined.html

Running claude-replay with no arguments opens a browser-based editor that auto-discovers your Claude Code and Cursor sessions. From there you can browse, edit, preview, and export replays visually.

For CLI usage, you can pass just a session ID — claude-replay will search ~/.claude/projects/, ~/.cursor/projects/, ~/.codex/sessions/, and ~/.gemini/tmp/ to find the matching file. Or pass the full path to a session file directly.

Cursor

Cursor transcripts are also supported — the format is auto-detected. Cursor transcripts don't include timestamps, so playback uses paced timing by default (see Timing modes).

claude-replay ~/.cursor/projects/*/agent-transcripts/<id>/<id>.jsonl -o replay.html

Codex CLI

Codex CLI (OpenAI) transcripts are also supported — the format is auto-detected. Codex tool calls (exec_command, apply_patch) are mapped to their Claude Code equivalents (Bash, Edit/Write) so they render with the same diff views and command previews.

claude-replay ~/.codex/sessions/2026/03/12/rollout-<id>.jsonl -o replay.html

Gemini CLI

Gemini CLI transcripts are also supported — the format is auto-detected. Gemini stores sessions as single JSON files (not JSONL) with inline thinking blocks and tool calls. Tool names are mapped to their Claude Code equivalents (run_shell_command → Bash, read_file → Read, etc.) for consistent rendering.

claude-replay ~/.gemini/tmp/<projectHash>/chats/session-<id>.json -o replay.html

You can also search by session ID:

claude-replay <session-id> -o replay.html  # searches ~/.gemini/tmp/ automatically

OpenCode

OpenCode transcripts are supported — the format is auto-detected. OpenCode stores sessions in a SQLite database, so you'll need to export the session first using the OpenCode CLI. Thinking/reasoning blocks are rendered natively.

# Export a session from OpenCode, then replay it
opencode export <sessionID> > session.jsonl
claude-replay session.jsonl -o replay.html

Kimi Code

Kimi Code (Moonshot AI) transcripts are supported — the format is auto-detected. Kimi Code stores sessions as wire.jsonl under ~/.kimi-code/sessions/, with each subagent in its own agents/<name>/ directory; the web editor discovers the main session alongside any subagents. Thinking blocks and tool calls are rendered natively.

Web Editor

The default experience. Launch it by running claude-replay with no arguments:

claude-replay
claude-replay --port 8080

The editor provides:

• Session browser — auto-discovers sessions from ~/.claude/projects/, ~/.cursor/projects/, ~/.codex/sessions/, and ~/.gemini/tmp/, plus a folder navigator for session files stored elsewhere
• Turn editor — include/exclude turns, edit user prompts, expand assistant blocks (read-only), add bookmarks
• Options panel — theme, speed, thinking/tool call toggles, redaction rules, labels
• Live preview — updates as you edit, renders the same output as the CLI
• Export — download the final HTML replay

The editor runs a local server on 127.0.0.1 (localhost only, not exposed to the network). It never modifies your original JSONL files — all edits are held in memory and only affect the exported output.

Usage

claude-replay [--port N]                        Launch the web editor (default)
claude-replay <input> [input2...] [options]     Generate replay from CLI
claude-replay extract <replay.html> [-o output.jsonl] [--format jsonl|json]

Each <input> can be a session file path or a session ID. If it is not an existing file path, it is treated as a session ID. claude-replay searches ~/.claude/projects/, ~/.cursor/projects/, ~/.codex/sessions/, and ~/.gemini/tmp/ for a matching session file. You can find your current session ID in Claude Code by running /status.

Multiple inputs are concatenated into a single replay (up to 20). When all sessions have timestamps, turns are sorted chronologically; otherwise command-line order is used. This is useful when accepting a plan creates a new session — chain the sessions to get the full story in one replay.

Commands

editor [file|session-id]

Launches the web-based replay editor. Optionally pass a file path or session ID to auto-load it on startup. See Web Editor above.

claude-replay editor                              # empty editor
claude-replay editor ~/.claude/projects/.../session.jsonl  # auto-load file
claude-replay editor abc123                       # auto-load by session ID

extract

Extract the embedded turn data from a previously generated replay HTML file. Outputs JSONL by default (one turn per line, bookmarks embedded). Use --format json for the legacy JSON format.

claude-replay extract replay.html -o session.jsonl            # JSONL (default)
claude-replay extract replay.html -o data.json --format json  # JSON

# Round-trip: extract, then regenerate with different options
claude-replay extract replay.html -o session.jsonl
claude-replay session.jsonl -o new-replay.html --theme dracula

The extracted JSONL can be fed back into claude-replay to regenerate with different options. Bookmarks are preserved as a bookmark field on each turn.

Options

Examples

# Replay turns 5 through 15 at 2x speed
claude-replay session.jsonl --turns 5-15 --speed 2.0 -o replay.html

# Filter by time range
claude-replay session.jsonl --from "2026-02-26T02:00" --to "2026-02-26T03:00" -o replay.html

# Clean output: no thinking, no tools
claude-replay session.jsonl --no-thinking --no-tool-calls -o replay.html

# Use a different theme
claude-replay session.jsonl --theme dracula -o replay.html

# Pipe to stdout for further processing
claude-replay session.jsonl --turns 1-5 > snippet.html

# Chain multiple sessions into one replay
claude-replay abc123 def456 ghi789 -o combined.html

# Live preview: serve and auto-reload on changes
claude-replay session.jsonl --serve --watch

# Serve without watching (one-shot preview)
claude-replay session.jsonl --serve

# Watch and write to file on changes
claude-replay session.jsonl --watch -o replay.html

Timing modes

The --timing flag controls how playback speed is derived:

paced mode creates presentation-style timing — similar to how slides appear in a presentation or subtitles are timed in a video. Block reveal speed scales with text length. This is the default for Cursor transcripts (which have no timestamps) and can also be used with Claude Code transcripts for smoother demos:

# Use paced timing even for Claude Code transcripts
claude-replay session.jsonl --timing paced -o demo.html

paced wording

The editor and CLI offer an additional text-reveal option inside paced timing. Choose paced wording in the editor, or pass --timing paced --pacing paced-wording on the CLI, to show prose at once in a dimmed state and then illuminate it one whole word at a time. The base pace is tunable from 80–600 WPM and defaults to 238 WPM, the estimated average adult silent reading rate for English non-fiction in Brysbaert's review and meta-analysis.

# Generate a paced-wording replay at 300 words per minute
claude-replay session.jsonl --timing paced --pacing paced-wording --reading-wpm 300 -o demo.html

The paced-wording mode varies the interval deterministically based on word length, adds bounded jitter, and pauses longer after punctuation and paragraph boundaries. It applies to user and assistant prose, while fenced code illuminates as one atomic block and thinking and tool blocks retain section-based reveal behavior. The regular speed control still scales playback live, and reduced-motion preferences disable word-by-word dimming in favor of the normal section-based reveal.

For the complete timing formula and tuning reference, see paced-word details.

Player controls

The generated HTML file is a fully self-contained interactive player:

• Play/Pause — auto-advances through turns with block-by-block animation
• Step forward/back — navigate one block at a time within turns
• Progress bar — click to jump to any point; session timer shows elapsed/total time
• Speed control — 0.5x to 5x playback speed
• Toggle checkboxes — show/hide thinking blocks and tool calls

Keyboard shortcuts

Themes

Built-in themes

claude-replay --list-themes

Available themes: tokyo-night (default), monokai, solarized-dark, github-light, dracula, bubbles.

Custom themes

Create a JSON file with CSS color values:

{
  "bg": "#0d1117",
  "bg-surface": "#161b22",
  "bg-hover": "#1c2128",
  "text": "#e6edf3",
  "text-dim": "#7d8590",
  "text-bright": "#ffffff",
  "accent": "#ff7b72",
  "accent-dim": "#c9514a",
  "green": "#3fb950",
  "blue": "#58a6ff",
  "orange": "#d29922",
  "red": "#f85149",
  "cyan": "#39d2c0",
  "border": "#30363d",
  "tool-bg": "#0d1117",
  "thinking-bg": "#0b0f14"
}

claude-replay session.jsonl --theme-file my-theme.json -o replay.html

Any missing keys are filled from the tokyo-night defaults, so you only need to specify the colors you want to change.

For advanced customization, add an extraCss key with arbitrary CSS rules to override layout, fonts, or any other styles:

{
  "bg": "#ffffff",
  "text": "#1c1e21",
  "extraCss": ".assistant-text { border-radius: 12px; border: 1px solid #ddd; }"
}

See the built-in bubbles theme for an example of a fully custom layout using extraCss.

Embedding

The output is a single HTML file with no external dependencies. Embed it in blog posts or docs with an iframe:

<iframe src="replay.html" width="100%" height="600" style="border: 1px solid #333; border-radius: 8px;"></iframe>

How it works

1. Parser reads the JSONL transcript line by line, handling Claude Code's streaming format (where a single assistant message appears as multiple lines with incremental content blocks)
2. Turns are grouped as: user message + assistant response (text, tool calls, thinking blocks) + tool results
3. Renderer compresses the parsed turns (deflate + base64) and injects them into the HTML template
4. The player is vanilla JS — no frameworks, no external requests. Data is decompressed at load time using the browser-native DecompressionStream API

Output optimization

Generated HTML files use two layers of optimization (zero external dependencies):

• Minified CSS/JS — the player template is minified with esbuild (mangled variable names, whitespace removed). Use --no-minify for readable output.
• Compressed data — transcript JSON is deflate-compressed and base64-encoded, typically reducing output size by ~60-70%. The browser decompresses it natively at load time using DecompressionStream (Chrome 80+, Firefox 113+, Safari 16.4+). For older browsers, use --no-compress to embed raw JSON.

Development

To rebuild the minified template after editing template/player.html:

npm install    # installs esbuild (devDependency)
npm run build  # generates template/player.min.html

The minified template is built in CI and included in npm releases. Without it, the CLI falls back to the unminified template automatically.

Secret redaction

By default, claude-replay scans all embedded text for common secret patterns and replaces them with [REDACTED] before they are written into the output HTML. This means secrets from your session (API keys, tokens, connection strings, etc.) never end up in the generated file.

Detected patterns include:

• API keys (sk-..., sk-ant-..., key-...)
• AWS access key IDs (AKIA...)
• Bearer and JWT tokens
• Database connection strings (postgres://..., mongodb://..., etc.)
• Private key blocks (-----BEGIN ... PRIVATE KEY-----)
• Generic key/value secrets (api_key=..., auth_token: ...)
• Environment variable secrets (PASSWORD=..., TOKEN=...)
• Long hex tokens (40+ characters)

Important: Pattern-based redaction is a best-effort safety net — it cannot catch every possible secret format. Always review the generated HTML before sharing publicly.

To disable redaction (e.g. for internal/private replays):

claude-replay session.jsonl --no-auto-redact -o replay.html

Supported transcript formats

Claude Code

One JSON object per line with a type field (user, assistant, system, progress, etc.). Includes timestamps, thinking blocks, and tool calls with results.

Cursor

One JSON object per line with a top-level role field. No timestamps. Thinking appears as inline text. The format is auto-detected — no flags needed.

Codex CLI

Event-based JSONL with typed events (session_meta, response_item, event_msg, etc.). Includes timestamps. Tool calls (exec_command, apply_patch) are mapped to Claude Code equivalents for consistent rendering. Codex's encrypted reasoning blocks are skipped; commentary messages are shown as thinking blocks. The format is auto-detected — no flags needed.

Gemini CLI

Single JSON file with a sessionId field and a messages array containing user and gemini typed entries. Includes timestamps, inline thinking blocks (thoughts), and tool calls with nested results. Tool names (run_shell_command, read_file, edit_file, etc.) are mapped to Claude Code equivalents for consistent rendering. The format is auto-detected — no flags needed.

Requirements

• Node.js 18+
• Zero runtime dependencies (esbuild is a dev-only dependency for building the minified template)

Privacy

Replay files embed the full session transcript, including source code, file paths, tool inputs/outputs, and thinking traces. Review the generated HTML before sharing publicly — it may contain proprietary code, internal paths, or other sensitive information. Secret redaction (enabled by default) catches common credential patterns but does not filter code or file contents.

The transcript data is stored as a compressed blob inside the HTML file. Editing the player JavaScript to hide or filter turns only affects rendering — the original data remains in the blob and can be recovered. To exclude sensitive content, use the CLI flags at generation time (e.g. --turns, --exclude-turns). Use --redact to strip specific strings (usernames, paths, project names) at generation time. Always review the generated replay before sharing publicly.

License

MIT