A local-first session analysis & review dashboard for your AI agents
"Hello Clawd, it's time for your weekly report"
-111827){kind=icon}
Timeline view — every session, traced |
AI session review — what you tried, what you learned |
Clawd Insights is a session recording & review dashboard for your agents. It automatically scans the work your local agents — Claude Code, Codex, OpenClaw, tclaude, Cursor and more — have already done, and turns it into a timeline with an AI-generated summary per session. Got tokens to spare? It can also digest every local agent conversation into your weekly work report.
Analysis data always stays on your machine. Session analysis runs through your own local claude / codex CLI (or another API backend you configure) — no third party ever sees it.
Currently macOS-first. Windows/Linux support is preliminary — issue reports are welcome. Requires Node.js v18 or later.
Option 1: command line
git clone https://github.com/yx0716/clawd-insights.git # grab the source
cd clawd-insights && npm install # install from source
npm start # run itOption 2: feed the prompt to your local agent, and let it do the install
Don't feel like wiring up the environment by hand? Paste this whole block into your local agent:
Install clawd-insights for me (a local agent session recording & review dashboard, repo: https://github.com/yx0716/clawd-insights):
1. First confirm git, node and npm are all available. If anything is missing, stop and tell me how to install it — don't improvise.
2. In a suitable directory (default ~/clawd-insights; if it already exists, ask me first), run:
git clone https://github.com/yx0716/clawd-insights.git && cd clawd-insights && npm install
3. Run npm start. Note it is a long-running desktop app — start it in the background; don't block in the foreground waiting for it to exit.
4. Once it starts, a pixel crab should appear on the desktop — seeing it means the install succeeded.
5. If any step fails, send me the exact error output and propose a fix — apply it only after I confirm.
6. Finally, tell me which directory it's installed in and how to start it next time.
Once it starts, a little crab appears on your desktop — right-click it and open the 📊 Session Review Dashboard to get going.
| Capability | What you get |
|---|---|
| Timeline view | Visualize every session by date / project / agent — a clear, global picture of what you've been doing |
| Calendar view | A calendar layout alongside the timeline, one click to switch |
| Local history scan | Reads local session records (Claude Code / Claude Internal / tclaude / Codex / Cursor / OpenClaw / opencode / Gemini CLI / Qwen Code) — nothing uploaded, no privacy risk |
| AI session review | Summarizes each conversation from your point of view: what you wanted solved, what the AI delivered, which tricks are worth reusing. Two depths — a quick digest, or a deep dive structured by the STAR method (Situation–Task–Action–Result) |
| Daily / weekly reports | One click at the top of the dashboard generates a daily or weekly report for the selected time range |
| Batch pre-analysis | Pre-generate summaries for recent sessions in bulk; provider-scoped caching makes them reusable |
| Cost tracking | Shows token usage and cost for every AI analysis |
① Open right-click → Dashboard |
② Pick a provider Local CLI / API / Ollama |
③ Tweak settings gear ⚙ → AI Provider |
④ Run analysis batch or per-session |
git clone https://github.com/yx0716/clawd-insights.git
cd clawd-insights && npm install
npm startOnce it launches, a small crab appears at the bottom-right of your desktop — it's the main entry point for session review.
There are three ways to open it — pick whichever feels natural:
- Right-click the desktop pet → choose 📊 Session Review Dashboard from the context menu
- Click the tray icon (menu bar on macOS) → 📊 Session Review Dashboard
- Keyboard shortcut: macOS
⌘ + Shift + Option + A
The session timeline works out of the box — nothing to configure. Smart session summaries, though, need a model behind them. There are three ways to provide one:
| Provider type | What it is | Setup | Best for |
|---|---|---|---|
| Local CLI (recommended) | Reuses the claude (Claude Code) or codex CLI already installed on your machine |
None — auto-detected | Anyone with a Claude Code / Codex / Cursor subscription — runs on the quota you already pay for, no extra cost |
| API key | An API key from Anthropic, OpenAI, etc. — pay-per-token | Paste the key into the dashboard settings | No local CLI installed, and happy to pay a little for analysis |
| Ollama | A locally-hosted open-model server (e.g. Ollama) | Point it at your local endpoint | Fully offline — never sends anything to the cloud |
💡 Strongly recommended: if Claude Code or Codex CLI is already on your machine, the dashboard auto-detects it and reuses your existing subscription quota — the simplest, most cost-effective option.
You can switch the analysis provider at any time in settings.
Once configured, you can analyze any session you want to review — or a specific time slice of one (useful when a session stretches over a long period with many breaks). Analysis comes in two depths: the quick version gives a digest overview, while the deep version breaks the work down with the STAR method (Situation–Task–Action–Result), ready to feed a project retrospective.
Every time you open the dashboard, if there are unanalyzed sessions, it automatically pops up a dialog — Pre-analyze Sessions — letting you analyze every session in a time range in one go.
Available scopes:
- Today — every session from today
- 3 Days — the last 3 days
- Week — the last 7 days
- Custom — your last N sessions
Pick a scope, hit confirm, and the dashboard shows an Analyzing 1/N, 2/N, ... progress bar while working through them in the background. Already-analyzed sessions are auto-skipped (per-provider cache), so re-clicking never wastes tokens.
If you only want to review one specific session, no batch needed — just click it:
- From the timeline — in Timeline view, click any colored block (each block is a session) and the detail card slides out on the right
- From the Sessions list — click any session card in the right-side list
Either way, the dashboard will:
- Show the cached summary first if it exists (sessions previously batch-analyzed are tagged
Analyzedand open instantly) - If not yet analyzed, the click immediately kicks off a single-session analysis — the card shows an
Analyzing…tag, and the result appears in seconds to tens of seconds
About cost: Local CLI (Claude Code / Codex subscription) analysis uses your existing subscription quota — typically no extra charges. In API key mode, the dashboard shows token usage and cost in the top status bar after each analysis completes, so you always know what you're spending.
If you skipped step 3, or want to switch providers later, adjust it via AI Provider Settings:
Open the dashboard → click the gear icon ⚙ in the top-right → the AI Provider Settings dialog appears.
- LOCAL CLI DETECTION — shows whether the dashboard found
claudeandcodexon your machine. Green dot = found (with version + path); red dot = missing. Green dots mean everything is working — carry on. - API PROVIDER (FALLBACK) — if no local CLI is installed, you can run AI session analysis with an API key (Claude / OpenAI / Ollama / …) — just paste the key.
Tip: if your
claudeorcodexwas installed via NVM, fnm, or Volta, auto-detection may miss it. Runwhich claudeorwhich codexin your terminal and paste the output into the Claude binary path / Codex binary path field.
- You have used
Claude Code,Codex, orCursor Agentlocally, and it still works - Local session history exists (it does by default)
Quick check
- Open Settings and look at
Local CLI Detection - Switch to
WeekorMonthand check whether the timeline shows any sessions
Clawd runs two independent data paths side by side:
Your Agent Clawd
│ │
├── live events ──→ hook / poll / plugin ──→ 🦀 pet animation
│ │
└── chat history ──→ local JSONL files ────→ 📊 session review dashboard
While an agent works (calling tools, waiting for input, erroring out, finishing a task…) it emits events. Clawd captures them through three integration modes and drives the pet accordingly:
| Mode | How it works | Latency | Agents |
|---|---|---|---|
| Command hook | Agent fires an event → automatically runs a script → script HTTP-POSTs the event to Clawd's local server (127.0.0.1:23333) |
Near zero | Claude Code, Claude Internal, tclaude, Copilot CLI, Gemini CLI, Cursor Agent, Kiro CLI, Antigravity CLI, CodeBuddy, Kimi CLI |
| Log polling | Clawd scans the agent's JSONL log file every ~1.5 s and detects new entries | ~1.5 s | Codex CLI, Gemini CLI (fallback), Kimi CLI (fallback) |
| In-process plugin | Plugin runs inside the agent's own runtime, forwarding events with zero overhead | Zero | opencode, openclaw, Hermes, Pi |
All events map to the same state machine: idle → thinking → working → happy / error → sleeping. The pet plays the matching SVG animation. When multiple sessions run simultaneously, it auto-switches to juggling / building / conducting animations.
Multi-agent coexistence: Claude Code, Claude Internal, tclaude, Codex, Copilot, Gemini, Cursor, Kiro, opencode, Antigravity CLI, CodeBuddy, Hermes, Kimi CLI, openclaw, and Pi can all run at the same time. Clawd tracks each session independently and displays the highest-priority state.
Every conversation you have with an agent is saved as JSONL on your disk:
| Agent | Local history path |
|---|---|
| Claude Code | ~/.claude/projects/ |
| Claude Internal | ~/.claude-internal/projects/ |
| tclaude | ~/.tclaude/projects/ |
| Codex CLI | ~/.codex/sessions/ |
| Cursor Agent | ~/.cursor/projects/ |
The insights dashboard reads these files directly to generate timelines and AI summaries. It doesn't go through hooks and doesn't require the pet to be running — as long as chat history exists on disk, the dashboard works.
Note: the analytics scanner currently covers only the four agents above. Copilot CLI, Gemini CLI, Kiro CLI, and opencode still drive pet animations, but their local histories are not yet wired into the dashboard scanner.
Beyond the analytics layer, this fork now tracks the full desktop-pet feature set from clawd-on-desk:
- Broader agent support — pixel-art reactions for Claude Code, Claude Internal, tclaude, Codex CLI, Copilot CLI, Gemini CLI, Cursor Agent, Kiro CLI, opencode, plus the newly merged Antigravity CLI, CodeBuddy, Hermes, Kimi CLI, openclaw, and Pi.
- WSL & remote development (focus of this sync) — Codex officially supports WSL2, and Clawd integrates through Codex's official hooks (with JSONL polling as a fallback). For an agent running on a remote box or inside WSL's separate Linux home, Remote SSH deploys the hooks over an SSH tunnel so the pet reacts to those sessions too. The remote side must use a POSIX shell — Git Bash or WSL
bash, not Windowscmd.exe. See docs/guides/codex-wsl-clarification.md and docs/guides/setup-guide.md. - GUI settings panel — a full settings window (agents, themes, shortcuts, Remote SSH, Telegram approval) replaces hand-editing config files.
- Theming — swap the crab for alternate characters (e.g. the Cloudling theme) or build your own with
npm run create-theme. - Telegram approval — approve or deny permission requests remotely from your phone.
- Doctor diagnostics — a built-in health check that verifies hook installation and per-agent integration status.
- Session HUD & Quick Commands — optional on-screen session-state labels and a quick-command surface.
- The classics — permission bubbles, mini mode, click reactions, eye-tracking, sleep sequences, and multi-monitor support all carried over.
Platform notes (Windows terminal focus, macOS focus, known limitations) live under docs/guides/.
Q: Does the dashboard need internet? Scanning and the timeline are fully offline. Whether AI summaries need internet depends on which provider you pick: Local CLI uses whatever network stack Claude Code / Codex normally use; Ollama is fully offline; API key mode talks to the cloud.
Q: Are my conversations uploaded anywhere? No. Clawd Insights collects zero telemetry. The provider step is your CLI or your API key calling the model you chose directly — no third-party server in the middle.
Q: I don't have Claude Code or Codex. Can I still use it? Yes. You can use the timeline view alone (completely free, no LLM required), or paste an Anthropic / OpenAI API key into AI Provider Settings to enable the cloud path.
Clawd Insights is a community-driven fork. Bug reports, feature ideas, and pull requests are all welcome — open an issue to discuss or submit a PR directly.
Credits are split in two: the people who build this fork (the analytics layer and everything listed above it), and the upstream community behind the desktop pet this fork stands on.
Thanks to everyone who has contributed to this fork:
 yx0716 |
 kingsley-wade |
 dopawei |
 XingLiu1 |
The desktop pet underneath — animations, permission bubbles, multi-agent tracking, themes, the lot — is the work of the clawd-on-desk community. The roll below is carried over from the upstream README (snapshot: July 2026); the canonical, always-current list lives there.
 @rullerzhou-afk 鹿鹿 · creator |
 @YOIMIYA66 maintainer |
 @Bynlk core contributor · Mobile / PWA |
Thanks to everyone who has helped make Clawd better:
Clawd Insights is the analytics layer built on top of rullerzhou-afk/clawd-on-desk, the lovely desktop pet that turns your coding agent's state into pixel art. Everything that makes the pet delightful — animations, permission bubbles, multi-agent state tracking, mini mode, the lot — is still here, untouched. What this fork adds is one extra question: what if every conversation you've ever had with the agent was searchable, summarised, and waiting for you on a single dashboard?
That dashboard is the new piece. It scans your local history (Claude Code, Claude Internal, tclaude, Codex CLI, Cursor Agent today; more agents on the way), draws a timeline, and lets your own LLM write per-session summaries — all without sending a byte to a third party.
Multi-agent state tracking carried over from upstream — and this repo now stays in sync with the full upstream codebase (WSL & Remote SSH support, the expanded agent roster, the GUI settings panel, theming, Telegram approval, doctor diagnostics, and more). For the desktop pet's own feature list, see Desktop pet capabilities above.
Huge thanks to @rullerzhou-afk and every contributor who shaped the original Clawd — this project wouldn't exist without that foundation. The full credit roll, fork and upstream side by side, is in Contributing above.
Source code: GNU AGPL-3.0-only — adopted from upstream clawd-on-desk, whose code this repo now incorporates. See also NOTICE.md for third-party materials.
Artwork (assets/) is NOT covered by the AGPL. All rights reserved by their respective copyright holders. See assets/LICENSE.
- Clawd character is the property of Anthropic. Unofficial fan project, not affiliated with Anthropic.
- Calico cat (三花猫) artwork by 鹿鹿 (@rullerzhou-afk). All rights reserved.
- Third-party contributions: copyright retained by respective artists.