llm-deep-trace

Browse, search, and analyze every conversation your agents have ever had.
All local. No accounts. No cloud.

What it is

You use coding agents. They leave behind session logs. Those logs are gold — decisions made, code written, reasoning traced, subagents spawned. But they're just JSONL files buried in ~/.claude/ or ~/.kimi/.

llm-deep-trace turns those files into a proper interface: threaded conversations, subagent trees, a conversation map, analytics, live tail, and full-text search across everything.

It runs entirely on your machine. It reads your local files. That's it.

Supported agents

Agent	Sessions	Subagents	Notes
Claude Code (Anthropic)	✓	✓	Full support incl. agent teams
Codex	✓	—
Kimi	✓	—	Think-block rendering
Gemini CLI (Google)	✓	—
OpenClaw	✓	✓
Cursor	✓	—	Partial — some metadata missing
Aider	✓	—	Chat history format
Continue.dev	✓	—
GitHub Copilot	✓	—
Factory Droid	✓	—
OpenCode	✓	—

Quick start

git clone https://github.com/userFRM/llm-deep-trace
cd llm-deep-trace
npm install
npm run dev

Open http://localhost:3000. No config needed — it finds your sessions automatically.

Requirements: Node.js 18+

Features

Session browser

• All providers in one sidebar, grouped and searchable
• Parent sessions show subagent count; click to expand the tree
• Expand all / collapse all, archive sessions you don't need

Conversation view

• User messages as bubbles, assistant responses as cards
• Tool call blocks collapsed by default, color-coded by type (bash, edit, web, browser, message, spawn)
• Thinking blocks (Claude extended thinking, Kimi reasoning) as expandable cards
• Inline image thumbnails with click-to-lightbox
• Ctrl+F find with match counter and highlights
• Skip preamble toggle — jump straight to the actual work

Subagent tree + conversation map

• React Flow visualization of the full session graph
• Click any node to navigate to that session
• Reset view, zoom capped at 75%, padding scales with graph size

Live tail

• Follow an active session in real time (3s poll)
• Green pulsing indicator while tailing

Analytics

• Sessions per day, stacked by provider
• Provider breakdown, top tools used, token totals
• Session length distribution

Search

• Full-text search across all sessions and all providers

Running remotely

llm-deep-trace reads files from the machine it runs on. If your agent sessions live on a remote server (VPS, home lab, cloud box), you have a few options:

SSH tunnel — forward the port to your local machine:

ssh -L 3000:localhost:3000 user@your-server
# then open http://localhost:3000 locally

Tailscale — start the server bound to your Tailscale IP:

npx next start --hostname <tailscale-ip> --port 3000
# accessible at http://<tailscale-ip>:3000 from any device on your Tailscale network

VPS / cloud — same as SSH tunnel, or bind to a private network interface. Do not expose port 3000 to the public internet — there's no auth.

In all cases: the app runs on the machine where your ~/.claude/, ~/.kimi/, etc. directories live.

How to ship it (for contributors)

Right now: clone and run locally. That's the intended workflow for a local devtool.

Planned:

• npm install -g llm-deep-trace + llm-deep-trace CLI command
• Homebrew tap for Mac users who prefer it
• Docker one-liner for the server crowd

PRs welcome on any of these.

Limitations

Being straight with you:

• Read-only. Browse and analyze — you can't resume or fork sessions from here yet. jazzyalex/agent-sessions does this for Codex and Claude Code.
• Local files only. Reads ~/.claude/, ~/.kimi/, etc. on the machine it runs on. Remote sessions not supported.
• Cursor / Aider / Continue support is partial. Session formats vary and some metadata is missing.
• No mobile UI. Built for desktop browsers.
• Requires Node.js 18+. Not a zero-dep install.

vs alternatives

jazzyalex/agent-sessions — the closest thing. Native macOS Swift app (macOS 14+), v2.11.2, actively maintained. Supports 7 agents: Codex, Claude Code, Gemini, Copilot, Factory Droid, OpenCode, OpenClaw. Has unified search, analytics, rate limits tracker, image browser, Homebrew install, and — notably — session resume (reopen a session directly in your terminal). If you're on a Mac and want a native app, it's worth looking at.

Where llm-deep-trace is different:

• Cross-platform — runs on Mac, Linux, Windows; anything with Node.js. Not macOS-only.
• More agents — 11 vs 7 (adds Kimi, Aider, Continue.dev, Cursor)
• Subagent tree — React Flow visualization of parent → child agent relationships
• Conversation map — full session graph, click-to-navigate
• Web-based — no install, no signing, no notarization; just npm run dev

What they have that we don't (yet):

• Session resume (reopen past sessions in the terminal)
• Real-time rate limits tracker
• Native performance + Homebrew cask + auto-updates

Tech stack

Next.js 16 · React 19 · TypeScript · React Flow · Zustand · highlight.js · Marked

License

MIT