Skip to content
100% local · zero telemetry · real-time everything

Real-time control plane for Claude Code.

A local-first dashboard that tracks every Claude Code session, agent, and tool call as it happens - and now spawns new ones for you. Sessions, kanban, analytics, workflows, MCP, plugin marketplace, and a 12-tab Config Explorer for everything Claude Code knows about.

Real-time hooks ingestion Stream-json output Conversation mode One-shot headless Resume sessions MCP server (25 tools) Plugin marketplace VS Code extension WebSocket push SQLite (WAL) OpenAPI 3.0 Mermaid diagrams macOS desktop app Native macOS toasts Hooks fan-out Tray live snapshot ⌘Q confirmation Auto-port discovery Discovery file CI auto-release Real-time hooks ingestion Stream-json output Conversation mode One-shot headless Resume sessions MCP server (25 tools) Plugin marketplace VS Code extension WebSocket push SQLite (WAL) OpenAPI 3.0 Mermaid diagrams macOS desktop app Native macOS toasts Hooks fan-out Tray live snapshot ⌘Q confirmation Auto-port discovery Discovery file CI auto-release
Compaction-aware tokens Per-model cost tracking Subagent hierarchy D3 Sankey workflows Run claude in-browser 12-tab Config Explorer Browser push notifications i18n (en / zh / vi) Activity feed Kanban board Update notifier Local-first, zero telemetry Run web + desktop together History import & backfill Native title bar Open-at-Login SMAppService integration Single-instance lock Per-arch DMG titles Compaction-aware tokens Per-model cost tracking Subagent hierarchy D3 Sankey workflows Run claude in-browser 12-tab Config Explorer Browser push notifications i18n (en / zh / vi) Activity feed Kanban board Update notifier Local-first, zero telemetry Run web + desktop together History import & backfill Native title bar Open-at-Login SMAppService integration Single-instance lock Per-arch DMG titles
Spawn Claude in one click Resume any session instantly Watch tool calls live Track cost per turn Replay full session history 60-second install Edit skills with backups Push notify on Stop Hooks fire in <5ms Single binary, single port Inspect every plugin Works fully offline One-DMG-per-build Quit cleanly w/ WAL checkpoint Auto-import legacy sessions PID-liveness pruning Dock icon stays on close Spawn Claude in one click Resume any session instantly Watch tool calls live Track cost per turn Replay full session history 60-second install Edit skills with backups Push notify on Stop Hooks fire in <5ms Single binary, single port Inspect every plugin Works fully offline One-DMG-per-build Quit cleanly w/ WAL checkpoint Auto-import legacy sessions PID-liveness pruning Dock icon stays on close
localhost:4820 - Analytics
Claude Code Agent Monitor - analytics page with token usage, tool frequency, activity heatmap and session trends
13
Pages
192
Server tests
148
Client tests
25
MCP tools
11
Workflow datasets
3
Languages (en/zh/vi)
The philosophy

Why local-first?

Your Claude Code data stays where it belongs - on your machine. No accounts to sign up for, no cloud service to trust, no telemetry quietly leaking back. Just a fast dashboard that watches what's already happening on your laptop.

🛡️

Zero telemetry

No accounts, no analytics, no third-party calls. The MCP server binds to localhost. Web Push needs your explicit opt-in. What you do here stays here.

Instant feedback

Hooks fire to the local server in milliseconds. WebSocket push updates the UI before you've taken your finger off the keyboard. No polling, no spinners.

📦

Yours forever

Everything's a SQLite file in ~/.dashboard/ and JSONL transcripts in ~/.claude/. Backup with cp, sync with whatever you already use, take it offline.

🎛️

Hooks-native

Built around Claude Code's own hook system. The dashboard never wraps, proxies, or monkey-patches your CLI. Your normal workflow continues even if the dashboard is down - hooks always exit zero.

What's inside

Everything you need, on your machine.

Hooks-driven session capture, deep analytics, in-browser Claude spawning, full configuration inspector, and a typed local MCP server with 25 tools.

Real-time dashboard

Six-tile overview, live agent cards with collapsible subagent hierarchy, recent activity feed, and a composite Health tab - all updated via WebSocket push, no polling.

Run Claude in-browser

Spawn claude subprocesses with a chat-style streaming UI. Conversation or one-shot mode. Resume any past session in one click. Slash + @-file autocomplete, live token meter, thinking-effort dial, and model picker.

🧰

Claude Config Explorer

12 tabs covering skills, subagents, slash commands, output styles, plugins, marketplaces, MCP servers, hooks, settings, memory, keybindings, statusline. Edit text-file artifacts with timestamped backups.

📊

Workflows visualization

D3-powered orchestration DAG, tool execution Sankey, collaboration network, error propagation map, concurrency timeline, complexity scatter, compaction impact - 11 datasets with rich i18n tooltips.

💰

Cost & token tracking

Per-model cost estimation with configurable pricing rules, compaction-aware token accounting that preserves totals across context compressions, and an LRU-cached JSONL extractor.

🔌

Local MCP server

25 typed tools across 6 domains. Three transport modes (stdio, HTTP+SSE, interactive REPL). Strict input schemas, retry/backoff, localhost-only API enforcement, and tiered safety gates.

🪝

Hooks-driven capture

Auto-installs into ~/.claude/settings.json. SessionStart, PreToolUse, PostToolUse, Stop, SubagentStop, Notification - every event lands in SQLite within milliseconds.

📦

Plugin marketplace

Official plugin marketplace with 5 plugins: ccam-analytics, ccam-productivity, ccam-devtools, ccam-insights, ccam-dashboard. 18 skills, 4 agents, 3 CLI tools, 2 hook configs.

🌐

Multi-language UI

English, Chinese, and Vietnamese - coverage extends end-to-end through every chart tooltip, Settings flow, Workflow narrative, Config Explorer tab, and Run page.

📋

Kanban board

Agents and sessions in toggleable views. Working / Waiting / Completed / Error / Abandoned columns with card cost, model, and current tool at a glance. WebSocket-scoped updates avoid unnecessary refetches.

📡

Swagger API docs

Full OpenAPI 3.0 spec with interactive Swagger UI. Try-it-out controls for every endpoint, request/response schemas, and collapsible endpoint groups - all served locally alongside the dashboard.

🧩

VS Code extension

Full dashboard sidebar inside VS Code. Monitor active sessions, view agents, and track tool calls without leaving your editor. Ships as a VSIX with zero external dependencies.

🍎

macOS desktop app

The whole dashboard as a native macOS .app, built with Electron 35. It embeds the Express server in-process, lives in the menu bar with a live status snapshot dropdown, auto-starts at login, and installs Claude Code hooks on first boot - zero-setup monitoring. Runs alongside npm run dev with hooks fanning out to both dashboards in parallel.

📰

Live activity feed

A real-time streaming event log with pause/resume buffering, multi-dimension filters, color-coded entries by event type, expandable hook payloads, and per-row session navigation. WebSocket-scoped updates so off-screen events never cost a refetch.

📥

History import & backfill

Rescan ~/.claude/projects/, point at any directory, or drag-drop a .tar.gz archive of transcripts. Imports are idempotent — per-event-type high-water marks dedupe re-runs — with WebSocket progress, archive extraction limits, and full subagent JSONL parsing.

🔔

Cross-surface notifications

Configurable per event type (new session, error, completion, subagent spawn). The web dashboard uses Web Push (VAPID); the macOS app fires native macOS toasts directly from the embedded server, side-stepping Web Push (which has no FCM credentials inside Electron). Both surfaces stay live when running together.

In action

See it for yourself.

Every page of the dashboard, from real-time monitoring to config editing, API docs, and even the VS Code extension. Click any screenshot to expand.

Dashboard overview with active agent cards and stats

Dashboard · Monitor

Overview stats, active agent cards, and recent activity feed

Dashboard system health tab

Dashboard · Health

Composite health score, storage donut, cache/error/success gauges

Analytics page with token usage and heatmap

Analytics

Token usage, tool frequency, activity heatmap, session trends

Sessions overview table

Sessions

Searchable, filterable, server-paginated table of recorded sessions

Session detail overview

Session Detail

Full session overview with metadata, cost, model, and tool breakdown

Session conversation tab

Session · Conversation

Live transcript with markdown, syntax-highlighted code, and styled tool calls

Session agents tab

Session · Agents

Overview tiles, subagent breakdown, and hierarchy tree

Session timeline tab

Session · Timeline

Chronological event timeline with multi-dimension filters

Kanban board agents view

Kanban Board · Agents

Working / Waiting / Completed / Error columns with live WebSocket updates

Kanban board sessions view

Kanban Board · Sessions

Sessions grouped by status across 5 columns

Activity feed

Activity Feed

Real-time event log with pause/resume and multi-dimension filters

Run Claude pre-run configuration

Run Claude · Config

Pick mode, source, cwd, model, permissions, and thinking effort

Run Claude live streaming output

Run Claude · Live Stream

Character-by-character streaming with collapsible tool calls and thinking blocks

Claude Config Explorer 12-tab inspector

Claude Config Explorer

12 tabs: plugins, skills, subagents, slash commands, MCP servers, hooks, and more

Workflows visualization

Workflows

D3-powered orchestration DAG, tool Sankey, collaboration network

Settings page

Settings

Model pricing rules, data management, and system information

Swagger OpenAPI documentation

Swagger API Docs

Interactive OpenAPI 3.0 spec with try-it-out controls

MCP Server interactive REPL

MCP Server

Interactive REPL exposing dashboard tools for Claude Code integration

Update checker modal

Update Checker

Commits-behind count with one-click copy-to-clipboard update command

Connection details modal with throughput sparkline and event-type breakdown

Connection Status

WebSocket details modal with 60-second throughput sparkline, top event types, and recent activity

Claude Code Monitor running as a native macOS desktop app

macOS Desktop App

The full dashboard, natively on macOS. Real window, embedded server, real-time intact — plus a menu-bar tray for quick access.

VS Code extension sidebar

VS Code Extension

Full dashboard sidebar inside VS Code with session monitoring

Import history settings panel

Import History

Rescan default folder, scan custom paths, upload archives with progress

Architecture

How the data flows.

A tight pipeline from CLI hooks to your browser. Everything's local - SQLite, your home dir, your machine. No accounts, no cloud, zero telemetry.

01

Claude Code fires a hook

Auto-installed entries in ~/.claude/settings.json trigger scripts/hook-handler.js on every SessionStart, PreToolUse, PostToolUse, Stop, SubagentStop, and Notification.

02

Hook handler POSTs

Stdin payload is parsed and POSTed to http://127.0.0.1:4820/api/hooks/event. Always exits zero - Claude Code never blocks on the dashboard being unreachable.

03

Server ingests & persists

Single SQLite transaction per event. Auto-creates sessions and agents. Extracts token usage, tool calls, compactions, API errors, and turn durations from JSONL transcripts via a stat-cached reader.

04

WebSocket push

session_*, agent_*, and new_event frames stream to every connected client. The React UI updates without polling. New run_stream / run_status frames power the Run page.

Built for

Who actually uses this.

Originally a tool for one person to keep track of one developer's Claude Code use. Turned out it's useful for anyone who ships AI-assisted code - solo or in a team.

01 - Solo dev

Watch what your CLI is doing

See every session, every tool call, every token spent - in real time. Cost tracking, compaction-aware accounting, kanban view of in-flight work.

Sessions Cost tracking Activity feed
02 - AI-ops engineer

Monitor team-wide CC usage

Deploy to Kubernetes with Helm, point hook handlers at the central instance, get Prometheus metrics + Grafana dashboards. Aggregate cost by project / model / user.

K8s Prometheus Multi-user
03 - Plugin author

Inspect & debug your plugin

Config Explorer shows exactly what your plugin contributes, how its hooks fire, and how its skills/agents are structured. Marketplace tab walks you through publishing.

Config Explorer Marketplace MCP server
04 - Power user

Run Claude from the browser

Spawn one-shot or multi-turn sessions, resume any past conversation, manage runs from a single page. Slash + @-file autocomplete, live token meter, effort dial, model picker, and permission mode all per-spawn.

Run page Resume Streaming Slash + @
05 - Indie hacker

Optimize your AI spend

Per-model pricing rules, daily / weekly / monthly cost trends, token-by-token analytics, compaction baselines so you know exactly what each session costs - even after a long thread compresses.

Pricing Analytics Cost
06 - Curious researcher

Study agentic workflows

D3 orchestration DAG, Sankey of tool transitions, collaboration network of subagent calls, error-propagation map - the workflow page is purpose-built for understanding how multi-agent CC sessions actually behave.

Workflows Subagents DAGs
07 - Team lead

Keep the squad aligned

See who's running what, which sessions are active across the team, and how many tokens each project consumes. Kanban board and activity feed keep standups honest.

Board Feed Multi-project
08 - VS Code native

Monitor without leaving the editor

Use the VS Code extension sidebar to watch sessions and tool calls while you code. No browser tab needed - everything docks into VS Code as a side panel.

VS Code Extension Sidebar
Quick start

Up and running in 60 seconds.

Local-first by design. Clone, install, run - your Claude Code sessions start showing up immediately because the dashboard auto-installs its hooks into your existing ~/.claude/settings.json.

Web app

Run it from your browser.

Four commands.

No accounts, no cloud, no telemetry. Everything stays on your machine. Existing sessions in ~/.claude/projects import on first start.

Full setup guide Browse code
~/Claude-Code-Agent-Monitor - zsh
# Clone & install
$git clone https://github.com/hoangsonww/Claude-Code-Agent-Monitor
$cd Claude-Code-Agent-Monitor
$npm run setup

# Install the hooks
$npm run install-hooks

# Start the dashboard (server + Vite)
$npm run dev

# Then open
http://localhost:4820
macOS app

Install it once, forget the terminal.

Prefer not to keep a terminal open? The dashboard also ships as a native macOS .app - an Electron 35 shell that embeds the same Express server in-process and renders the built React client in a native window. Same dashboard, macOS-native lifecycle on top.

01

Native, menu-bar-first

A menu-bar (tray) icon with a single-click dropdown showing a live status snapshot (sessions, agents, events today), a native macOS title bar with the conventional traffic-light row, single-instance lock, and window-close-hides-but-server-keeps-running. ⌘Q opens a confirmation dialog (press ⌘Q again to skip it).

02

Auto-start at login

Toggle Open at Login from the tray menu - it registers through macOS's native Login Items, so the dashboard is always running when you sit down.

03

Zero-setup, durable data

On first boot the app installs Claude Code hooks and starts background services automatically - a DMG-only user gets monitoring with essentially no manual setup. Your SQLite database lives in ~/Library/Application Support/Claude Code Monitor/data/, outside the .app bundle, so it survives reinstalls and updates. The app also recovers your login-shell PATH at startup, so Run Claude can find and launch the claude CLI. Notifications are delivered as native macOS toasts from the embedded server (Web Push doesn't work reliably inside Electron).

04

Coexists with the web dashboard

Run the desktop app and npm run dev at the same time and both stay real-time. Every running dashboard appends its {port, pid, startedAt} entry to ~/.claude/.agent-dashboard.json on startup; the hook handler fans each hook out to every live entry in parallel, so both UIs receive the same events and both their SQLite stores stay in sync. Stale entries self-evict via a PID liveness check on read.

Install steps

From DMG to running in under a minute.

The DMG isn't notarized through Apple (yet), so the first launch needs one extra Terminal command to clear macOS's quarantine flag. Everything after that is normal macOS lifecycle.

01

Download the right DMG

Open the latest release and grab ClaudeCodeMonitor-<version>-arm64.dmg for Apple Silicon (M1 / M2 / M3 / M4) or -x64.dmg for Intel Macs. Not sure which? Open the Apple menu → About This Mac: anything labelled "Apple" silicon is arm64.

02

Drag into Applications

Double-click the DMG to mount it, then drag Claude Code Monitor.app onto the Applications shortcut beside it. Eject the DMG once the copy is done.

03

Clear the Gatekeeper quarantine

In Terminal, run: xattr -cr "/Applications/Claude Code Monitor.app". Without this, macOS shows "Apple could not verify…" because the build is ad-hoc signed, not notarized. One-time step per install.

04

Launch & toggle auto-start

Open Claude Code Monitor from Applications (or Spotlight). The menu-bar icon appears within a second; click it → Open at Login and macOS will start the app at every login from now on - managed through System Settings → General → Login Items.

Get it in two ways.

Download a pre-built DMG - no build required. Hit the button below to land on the latest GitHub Release; grab the matching arch DMG - ClaudeCodeMonitor-<version>-arm64.dmg for Apple Silicon or -x64.dmg for Intel - from its assets and drag the app into /Applications. The link always redirects to the current release because CI publishes a new vX.Y.Z automatically every time the version in package.json is bumped on master. Releases are public - no GitHub sign-in required.

Want a per-commit build from CI instead of waiting for a release? The macOS Desktop (DMG) workflow uploads a ClaudeCodeMonitor-dmg artifact on every green run - useful for testing master before it's tagged. Or build it yourself: a single-arch build (desktop:dmg:arm64 or desktop:dmg:x64) finishes in about a minute, while desktop:dmg produces both DMGs in one go - the same dual output the release workflow ships - and is correspondingly slower. Each DMG is roughly 125 MB, about 250 MB installed.

Download the latest release Per-commit CI builds Full desktop app guide Architecture reference
~/Claude-Code-Agent-Monitor - zsh
# Build the macOS .app locally
$npm run setup
$npm run build
$npm run desktop:install

# Fast - one architecture (recommended for your own Mac)
$npm run desktop:dmg:arm64   # Apple Silicon
$npm run desktop:dmg:x64     # Intel

# Both arches - two separate DMGs, what the release workflow ships
$npm run desktop:dmg

# Open the DMG, drag the app to Applications, then clear
# the Gatekeeper quarantine on first launch:
$xattr -cr "/Applications/Claude Code Monitor.app"
Tour the dashboard

A purpose-built page for every job.

Each route is its own focused workspace - kanban for at-a-glance status, sessions for drill-down, analytics for trends, run for live spawning. They all sit on the same hook-driven event store, so what you see anywhere is the same source of truth.

Built with

A focused stack.

Boring, battle-tested foundations. No SaaS, no databases-as-a-service, no authentication providers. Just your machine + the tools that actually matter.

⚛️
React 18
+ TypeScript 5
Vite 6
HMR + bundle
🚂
Express 4
Node.js 18+
🗃️
SQLite
better-sqlite3 + WAL
🔌
WebSocket
RFC 6455 push
📡
MCP 1.0
stdio / HTTP / REPL
🎨
Tailwind 3
+ lucide icons
📊
D3.js 7
Sankey + DAG
🌐
i18next
en / zh / vi
📋
OpenAPI 3.0
+ Swagger UI
🔔
Web Push
VAPID
🐳
Docker
+ Podman
☸️
Kubernetes
+ Kustomize + Helm
🌿
Terraform
AWS / GCP / Azure / OCI
🔥
Prometheus
+ Grafana
📈
Mermaid
10.x diagrams
🧪
Vitest
+ Testing Library
Node test
node:test runner
🪝
CC hooks
native CLI integration
🔧
React Router
v6 client routing
🎯
i18n detector
browser language
📦
multer
multipart upload
🗜️
tar / zip
safe archive extract
🔍
ESLint
+ Prettier
🎬
PostCSS
+ Autoprefixer
🔄
workbox
service worker (planned)
🐙
GitHub Actions
CI / release
🦊
GitLab CI
pipelines
📦
VS Code
extension API
🐍
Python 3
statusline script
☁️
Coralogix
observability
🔭
OpenTelemetry
collector
🖥️
Electron 35
macOS .app shell
📦
electron-builder
DMG packaging
🤖
Claude CLI
spawned via Run page
🔔
web-push
VAPID push library
📄
swagger-ui-express
interactive API docs
📤
multer
multipart upload
🗜️
adm-zip
zip extract (bombs-safe)
🪢
tar
tar/tgz extract
🌍
cors
Express CORS
🧵
ws
WebSocket server
🔑
SMAppService
macOS Login Items
🚦
concurrently
dev process orchestrator
⚙️
@types/node
strict TS in desktop/
🆔
uuid
v4 ids
🪟
Chromium
embedded BrowserWindow
🍎
macOS native APIs
tray, dock, dialog
FAQ

Frequently asked questions.

Quick answers to the things people ask most. The wiki has the full picture.

Does this send any data to the cloud?

No. The dashboard is a 100% local-first Node.js + React app that reads your existing ~/.claude/ directory and writes to a local SQLite file. There is zero telemetry, zero analytics, no third-party calls. The MCP server is bound to localhost. Web Push notifications use VAPID and only push to subscribers who explicitly opted in via your own browser.

How does it integrate with Claude Code?

Claude Code's native hook system. npm run setup installs hook entries into ~/.claude/settings.json that POST every SessionStart, PreToolUse, PostToolUse, Stop, SubagentStop, and Notification event to the local server. Hooks always exit zero so they never block your CLI; if the dashboard is down, your workflow is unaffected and events resume capture as soon as it's back.

Will running Claude from the dashboard cost extra tokens?

The Run page spawns real claude processes - they consume your normal Claude Code quota / API tokens, the same as if you ran them in your terminal. The dashboard adds no overhead. The cost-tracking analytics show you exactly what each run consumes, by model, with compaction-aware accounting.

Can I resume any session, even ones I didn't start in the dashboard?

Yes. The Run page's "Resume existing session" picker lets you select any session from your full history (everything that's ever lived in ~/.claude/projects/), spawn claude --resume <id> with the original cwd, and continue the conversation from where it left off - directly in your browser.

Does the Run page have feature parity with the terminal TUI?

Most of what matters, with the limits the stream-json interface forces on us. We support slash-command autocomplete (user / project / plugin commands expanded client-side, plus built-in CLI commands flagged "CLI only - won't run from here"), @-file references with debounced cwd-scoped fuzzy search, a live context-window / token meter showing input + output + cache-read with cost, a thinking-effort field wired to --effort, multi-turn conversation, session resume, real character-by-character streaming, and full hook capture. What doesn't carry over is the TUI itself - in-place tool approval prompts, interactive /login / /init, the keybinding overlay - those are CLI-only. There's a collapsible limitations banner on the Run page that spells out exactly which is which, and any session can be popped open in your terminal with claude --resume <id>.

Is it safe to edit my skills / agents from the Config Explorer?

Yes. Every mutation creates a timestamped backup in <CLAUDE_HOME>/cc-config-backups/<type>/ before writing. Backups land outside the directories Claude Code scans, so a deleted skill cannot resurface as a backup-named one. Writes are atomic via temp file + rename. Plugins, MCP servers, hooks-in-settings, and live settings.json files are intentionally read-only because they have concurrent-write races with the running CLI.

Is there a desktop app, and how is it different from running the server?

Yes - a native macOS .app built with Electron 35. It embeds the same Express server in-process and renders the built React client in a native window, so it is the same dashboard with macOS-native lifecycle on top: a menu-bar icon, a native app menu, auto-start at login, a single-instance lock, and window-close-hides-but-server-keeps-running. It also installs Claude Code hooks and starts background services on first boot, so you get monitoring with essentially zero manual setup. Get it by downloading the ClaudeCodeMonitor-dmg CI artifact or building locally with npm run desktop:dmg:arm64 / :x64 (fast) or npm run desktop:dmg (universal, slow). The DMG is ad-hoc signed, so clear Gatekeeper on first launch with xattr -cr "/Applications/Claude Code Monitor.app". Your SQLite database lives in ~/Library/Application Support/Claude Code Monitor/data/ outside the .app bundle, so it survives reinstalls and updates, and the app recovers your shell PATH at startup so Run Claude can launch the claude CLI. macOS only for now - see the MacOS app guide.

What about scaling / multi-user?

The default deployment mode is single-user, localhost-bound. For team setups we ship Docker / Podman compose files, Kubernetes manifests with Kustomize, Terraform modules for AWS/GCP/Azure/OCI, Helm charts, and Prometheus/Grafana observability - see the deployment guide.

One-line install · zero config

Take Claude Code from black box to glass box.

See every session, agent, and tool call as it happens. Spawn and resume Claude from your browser. Inspect every plugin, skill, hook, and setting in 12 tabs. All on your machine - no telemetry, no signup, no cloud.

MIT licensed 100% local-first Zero telemetry 13 pages, 1 binary Works offline