LLMBoard
dcreddi/llmboard
Dashboard em tempo real para monitorar agentes Claude, ferramentas e segurança
Ver no GitHubDashboard de observabilidade em tempo real para Claude Code que monitora agentes, chamadas de ferramentas, tokens e anomalias de segurança. Fornece visibilidade completa sobre sessões, projetos, comandos executados e atividade de rede — tudo acessível via navegador local.
Instalação
npx convez-ai dcreddi/llmboard1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Monitorar sessões ativas de agentes Claude e consumo de tokens
- Detectar anomalias de segurança e comandos perigosos em tempo real
- Analisar padrões de uso de ferramentas e domínios externos contatados
- Gerenciar subagentes paralelos e inspecionar timeline de execução
Veredito

LLMBoard preenche uma lacuna crítica: observabilidade nativa e local para agentes Claude. A implementação é sólida, com captura automática de eventos, UI responsiva e foco em segurança — essencial para quem trabalha com Code em produção.
README
A tradução para o português deste README ainda está na fila.
LLMBoard
Real-time observability dashboard for Claude Code. Watch agents, tool calls, tokens, commands, and security anomalies as they happen — all in your browser at localhost:3456.
Quick Start
npm install -g llmboard
llmboard setup
llmboard
That's it. Your browser opens to http://localhost:3456.
What It Shows
| Tab | What you see |
|---|---|
| Live Sessions | Every active/stopped session — model, path, tool calls, tokens, duration. Click any row to expand the tool timeline, git diff, and CLAUDE.md rules. |
| Project Intelligence | All projects detected from session paths — git branch, remote URL, git identity, last commit, session count, alert summary. |
| Agent Manager | Parallel subagents with status, task description, and duration |
| Anomaly Alerts | Security alerts: dangerous commands, sensitive data detected, prompt injection risk, new external domains |
| Command Center | All Bash commands classified as Safe / Needs Info / Needs Permission / Dangerous, plus a custom rules editor |
| Tool Activity | Real-time feed of every tool call across all sessions |
| Token Analytics | Total/input/output tokens, tokens over time, tool usage distribution, daily breakdown |
| Network & Domains | Every external domain contacted — categorized, with call counts and session links |
| Settings | Theme, default model, token estimation, data export, bug reporting |
Session row anatomy
› ACTIVE Sonnet ACCEPTEDITS .../myapp 42 tools 12.4K tok 03:21
| Element | Meaning |
|---|---|
› | Click to expand — shows tool timeline, git diff, CLAUDE.md |
ACTIVE / STOPPED | Live session state |
Sonnet | Model auto-detected from the event stream |
ACCEPTEDITS | Permission mode |
| Path | Working directory (truncated, full path in tooltip) |
42 tools | Tool calls so far this session |
12.4K tok | Estimated tokens (input + output) |
03:21 | Elapsed time |
Click the Ext. Domains or Alerts counters in the header to jump directly to those tabs.
Installation
Requirements
- Node.js 18+ (cross-platform — macOS, Linux, Windows)
jqis optional and only used by the legacy shell hook; the default hook is a zero-dependency Node script.
Option A — Claude Code plugin (recommended)
Installs the event hooks automatically without modifying ~/.claude/settings.json, and works the same on every platform. In Claude Code:
/plugin marketplace add dcreddi/llmboard
/plugin install llmboard@llmboard
Then launch the dashboard any time with:
/llmboard:launch-dashboard
(or run npx llmboard in a terminal). The plugin captures events to ~/.llmboard/; the dashboard reads from there.
Use either the plugin or
llmboard setup(Option B) — not both, or events are logged twice.
Option B — npm package
npm install -g llmboard
llmboard setup
llmboard setup is explicit (no install-time side effects). It will:
- Create the
~/.llmboard/data directory and default config - Register event hooks in
~/.claude/settings.json(idempotent — safe to re-run) - Offer to install a background auto-start service (launchd / systemd / Task Scheduler)
Run npx llmboard (no install) to start the dashboard, or llmboard once installed globally. Uninstall cleanly with llmboard remove (add --purge to also delete ~/.llmboard/ data).
CLI Reference
llmboard Start dashboard + open browser
llmboard setup Full setup (hooks + service)
llmboard remove Uninstall hooks, restore settings.json
llmboard service install Install auto-start service (runs on login)
llmboard service uninstall Remove auto-start service
llmboard service status Show service status
llmboard stats Print token/session summary to terminal
llmboard stats --today Today's breakdown
llmboard stats --week Weekly breakdown
llmboard export --csv Export analytics as CSV
llmboard export --json Export analytics as JSON
llmboard config set <key> <val> Update config
llmboard doctor Diagnose hooks, jq, port, event flow
Options:
--port <n> Custom port (default: 3456)
--no-open Don't auto-open browser
--share Share via Cloudflare Tunnel (requires cloudflared)
VS Code Extension
The vscode-extension/ directory contains a companion extension that embeds the dashboard as a panel inside VS Code.
cd vscode-extension
npm install
npm run package
code --install-extension llmboard-1.0.0.vsix
The extension:
- Starts the LLMBoard server automatically when VS Code opens
- Shows a status bar item (
⬡ LLMBoard) — click to open the panel - Command palette:
LLMBoard: Open Dashboard/LLMBoard: Stop Server
Architecture
Claude Code
│
▼
hooks/event-logger.sh ← bash script, ~10ms, never blocks
│ writes JSONL
▼
~/.llmboard/events.jsonl ← append-only log, rotates at 50MB
│ chokidar file watch
▼
src/server/ ← Express + WebSocket server
file-tailer.js ← byte-offset tailing, handles rotation
event-store.js ← in-memory index (sessions, domains, anomalies)
routes.js ← REST API (/api/sessions, /api/events, etc.)
index.js ← WS broadcast + ping/pong heartbeat
│ WebSocket
▼
src/public/ ← Vanilla JS dashboard (no build step)
dashboard.js ← WS client, view router, event bus
components/ ← One file per tab
Key design decisions:
- Shell script hook — ~10ms vs ~100ms for a Node.js process; hooks fire on every tool call
- JSONL file — simple append from bash; server builds the in-memory index
- Chokidar — reliable file watching on macOS (fs.watch misses events)
- Vanilla JS — no build step, no bundler, <50KB total
- Zero external calls — all data stays on your machine
Data & Privacy
All data is stored locally in ~/.llmboard/:
| File | Purpose |
|---|---|
events.jsonl | Raw event stream — auto-rotated at 50MB |
config.json | User preferences (port, model, theme) |
settings-backup.*.json | Pre-setup backup of ~/.claude/settings.json |
service.log | Background service stdout/stderr |
Zero telemetry. Zero external calls. Nothing leaves your machine.
Configuration
Edit ~/.llmboard/config.json or use llmboard config set <key> <value>:
{
"port": 3456,
"model": "sonnet-4",
"retention_days": 30,
"max_file_size_mb": 50,
"auto_open_browser": true,
"theme": "dark"
}
How-To Guides
Install LLMBoard as a PWA (desktop app)
Open http://localhost:3456 in Chrome or Edge and click the install icon in the address bar. LLMBoard installs as a standalone desktop app — no browser chrome, opens from your dock.
Change the dashboard port
llmboard config set port 4000
llmboard # starts on 4000
# or one-off:
llmboard --port 4000
Update any bookmarks or VS Code extension config to match.
Share the dashboard with a teammate
llmboard --share # requires cloudflared to be installed
Creates a temporary public Cloudflare Tunnel URL valid for the session. No data is stored remotely — the tunnel proxies your local server only.
Export data for analysis
llmboard export --csv > sessions.csv
llmboard export --json > sessions.json
Fields exported: session_id, cwd, started_at, ended_at, status, tool_calls, input_tokens, output_tokens, model.
Run LLMBoard as a background service
llmboard service install # launchd (macOS) / systemd (Linux) / Task Scheduler (Windows)
llmboard service status
llmboard service uninstall
After install, the dashboard starts automatically at login.
View session activity in detail
Click any session row (› chevron) to expand:
- Tool timeline — every tool call in order with inputs/outputs
- Git diff — uncommitted changes in the session's working directory at session time
- CLAUDE.md — the active CLAUDE.md rules that governed the session
Add a custom command rule
Open Command Center → scroll to Custom Rules → add a regex pattern and choose a classification (Safe / Needs Info / Needs Permission / Dangerous). Rules apply immediately to all future commands.
View git identity per project
Open Project Intelligence. Each project card shows:
- Remote — the
originremote URL - Git user / email — the local
git configidentity for that repo - Last commit — short hash, subject, and relative time
Useful for confirming which GitHub account is active on each project.
Diagnose a broken setup
llmboard doctor
Checks: Node ≥ 18, jq available, hooks in ~/.claude/settings.json, events file writable, port 3456 free, WebSocket connects.
Acknowledge an anomaly alert
Open Anomaly Alerts → click Acknowledge on any alert. Acknowledged alerts are dimmed and won't re-trigger for the same event.
Query the REST API directly
The server exposes a full REST API at http://localhost:3456:
| Endpoint | Description |
|---|---|
GET /api/sessions | All sessions with stats |
GET /api/events?limit=50 | Recent raw events |
GET /api/stats | Global token + session totals |
GET /api/export?format=csv | CSV export |
GET /api/network | External domains seen |
GET /api/anomalies | All anomaly alerts |
GET /api/projects | Project summaries |
GET /api/digest | Daily digest (sessions, tokens, top tools) |
GET /api/skills | Skill/plugin registry |
GET /api/git?cwd=<path> | Git info for any local directory |
GET /api/sessions/:id/git | Git diff for a specific session |
GET /api/sessions/:id/claudemd | CLAUDE.md rules for a session |
GET /api/commands | All logged Bash commands |
GET /api/command-rules | Custom command classification rules |
GET /api/health | Server health check |
Training Guide
Understanding the dashboard tabs
| Tab | When to use it |
|---|---|
| Live Sessions | See what agents are doing right now — click to expand tool timeline, git diff, CLAUDE.md |
| Project Intelligence | Cross-session view per repo: sessions, identity, branch, alerts, last commit |
| Agent Manager | Debug parallel subagents — which are running, how long, what task |
| Tool Activity | Audit every file read, Bash command, and web fetch in real time |
| Command Center | Review shell commands with safety classifications; add custom rules |
| Token Analytics | Track usage trends, find expensive sessions, plan budget |
| Network & Domains | Every external domain contacted — useful for security reviews |
| Anomaly Alerts | Dangerous commands, sensitive data, prompt injection, new domains |
| Settings | Theme, model for token estimates, data export, feedback links |
Understanding anomaly types
| Alert type | What triggered it |
|---|---|
| Dangerous command | Bash matched a high-risk pattern (rm -rf, curl | bash, chmod 777, etc.) |
| Sensitive data | Tool input/output contained API keys, passwords, or tokens |
| Prompt injection risk | Tool result contained suspicious instruction-like content |
| New external domain | A domain contacted for the first time in this session |
Command safety classifications
Every Bash command is automatically classified:
| Label | Meaning |
|---|---|
| Safe | Read-only — ls, cat, grep, git status |
| Needs Info | Modifies files or state |
| Needs Permission | Installs, network calls, service changes |
| Dangerous | High-risk — rm -rf, chmod 777, curl pipe, ngrok |
Add your own rules in Command Center → Custom Rules.
Reading Project Intelligence
Each project card shows all sessions Claude ran in that directory, the git branch and remote, and the git identity (user + email) configured locally. If a project shows a different git email than expected, it means the repo has a local git config user.email override — LLMBoard reads the effective config, not just the global one.
FAQ
Q: Does LLMBoard slow down Claude Code? The hook is a shell script (~10ms). Claude Code's tool calls take hundreds of milliseconds minimum — LLMBoard adds no perceptible delay.
Q: Does my data leave my machine?
No. Everything stays in ~/.llmboard/events.jsonl served from a local Express server. The only exception is llmboard --share, which proxies via Cloudflare Tunnel — only if you explicitly pass --share.
Q: Why are token counts estimates? Claude Code hooks don't expose the actual API token counts. LLMBoard estimates from character length (~4 chars/token). Estimates are labeled clearly in the UI.
Q: The dashboard shows "No Active Sessions" even though Claude is running.
Run llmboard doctor. Common causes: hooks not installed (llmboard setup), ~/.llmboard/events.jsonl not writable, or the server started before any events were logged.
Q: How do I reset all data?
rm ~/.llmboard/events.jsonl && touch ~/.llmboard/events.jsonl
The server picks up the empty file immediately — no restart needed.
Q: Can I run multiple dashboards on different ports?
Yes: llmboard --port 3457 --no-open in a second terminal. Both watch the same events.jsonl.
Q: What does the VS Code extension add?
It embeds the dashboard as a side panel so you don't need a separate browser tab and auto-starts the server when VS Code opens. See vscode-extension/.
Q: How do I add my team's internal domains to the allowlist? There's no allowlist UI yet — all external domains appear in Network & Domains. You can filter by session there. Domain allowlisting is on the roadmap.
Q: Why does Project Intelligence show the wrong git identity?
LLMBoard reads the effective git config user.name and user.email for each repo (local config takes priority over global). If a project shows unexpected credentials, check git config --list in that directory.
Q: How do I skip the automatic setup on install? Set the environment variable before installing:
LLMBOARD_SKIP_SETUP=1 npm install -g llmboard
Then run llmboard setup manually when ready. Setup is also skipped automatically in CI environments (CI=true).
Q: I'm getting 429 Too Many Requests from the API. The server rate-limits API calls to 120 requests/minute per client to prevent runaway scripts from overwhelming it. If you're hitting this legitimately (e.g. a custom integration), add a small delay between calls or batch your requests.
Q: Can I use LLMBoard without installing it globally? Yes:
git clone https://github.com/dcreddi/llmboard
cd llmboard
npm install
node bin/cli.js setup
node bin/cli.js
Q: How do I uninstall completely?
llmboard remove # remove hooks from ~/.claude/settings.json
llmboard service uninstall # remove auto-start service
npm uninstall -g llmboard
rm -rf ~/.llmboard # delete all local data
Contributing & Feedback
Found a bug or have an idea?
| Type | Link |
|---|---|
| Bug report | Open an issue |
| Feature request | Request a feature |
| Questions | GitHub Discussions |
You can also click Report a Bug or Request a Feature directly from the Settings tab inside the dashboard.
What to include in a bug report
- LLMBoard version —
llmboard --version - OS and Node version —
node --version,uname -a - Doctor output —
llmboard doctor(paste full output) - What you expected vs what happened
- Steps to reproduce
For event pipeline issues, also include:
tail -5 ~/.llmboard/events.jsonl
License
MIT — Copyright (c) 2026 DCReddi Inc (katuru@dcreddi.com)