open source · MIT · npm

See Inside Your AI Agent's Mind.

Live Kanban · Session Analytics · Agent API
for Claude Code agents. Zero setup, no database.

View on GitHub ↗
~ npx -y claudedash@latest start → localhost:4317 ▋
Core Features

Full-stack observability
for AI agents.

Live Kanban Board

Real-time task status from ~/.claude/tasks/. Watch your agent's todo list update as it works — Plan, In Progress, Review, Done.

SSE updates TodoWrite zero polling

Plan Mode & Dependencies

Parse queue.md with task dependencies, acceptance criteria, and blocked status. Full dependency graph visualization.

queue.md execution.log dep graph

Quality Gates

Log lint, typecheck, and test results per task via meta.quality. See ✅/❌ inline in each Kanban card with full timeline history.

lint typecheck tests

Context Health Monitoring

Token-based context window estimation. Green → yellow → red warnings when sessions approach 65% and 75% usage. Aggregate banner in live mode.

65% warn 75% critical per session

Session Quality & AI Facets

Reads ~/.claude/usage-data/facets/ — AI-generated session analysis with outcome scores, helpfulness ratings, friction types, and satisfaction rates.

outcome helpfulness satisfaction

Agent API

HTTP endpoints for multi-agent workflows: POST /log to record task results, GET /queue for computed statuses, agent registry with heartbeat & stale detection. BLOCKED events push browser notifications instantly.

POST /log GET /queue agent registry

Hook Architecture & Cost Tracker

Install Claude Code lifecycle hooks with claudedash hooks install. PostToolUse events stream via SSE. 5-hour rolling billing block estimates cost in real time across all models.

PostToolUse billing block per-model cost
Dashboard

Everything in one view.

Session sidebar, Kanban board, context health bars, quality gates — all updating via SSE.
Zero refresh. No database. No cloud.

🔒 localhost:4317
claudedash Live view — session Kanban with context health bars claudedash Kanban view — tasks across PENDING, IN PROGRESS, COMPLETED columns claudedash Queue view — plan mode task board with dependency graph claudedash Context Health — live token usage per session with color-coded warnings claudedash Worktrees view — parallel agents across git branches claudedash terminal startup — npx claudedash start animated demo
Updates in real-time via SSE — typically within ~100ms of every file write
Quick Start

Up in 30 seconds.

Terminal 
# Zero-install — always gets latest version
$ npx -y claudedash@latest start

# Install real-time lifecycle hooks
$ claudedash hooks install
✓ Hooks installed in ~/.claude/settings.json

# Set up plan mode in your project
$ claudedash init
✓ Created .claudedash/ with queue.md, workflow.md

▶ Live sessions · Queue tracker · Agent registry
✓ Server running at http://localhost:4317
01

Run the CLI

Fastify server starts at localhost:4317. Chokidar watches ~/.claude/ and .claudedash/ for file changes via SSE.

02

Install hooks (optional)

Run claudedash hooks install to inject PostToolUse/Stop hooks into Claude Code. Every tool call streams live to the dashboard event feed.

03

Agents log via HTTP

Agents call POST /log to record task results. BLOCKED status triggers instant browser notifications. GET /queue returns dependency-resolved task statuses.

Git Worktrees & Analytics

Multi-repo agent visibility.

Running agents across multiple git worktrees? claudedash maps sessions to branches by cwd, shows dirty state and ahead/behind counts — and the Activity tab surfaces tool usage, prompt history, cost breakdown, and AI-scored session quality.

git worktree list dirty detection ahead / behind task mapping tool analytics cost tracker AI quality scores
claudedash Worktrees tab — parallel agents across git branches with context health
Claude Code — MCP query
❯ claude mcp add claudedash -- npx -y claudedash@latest mcp
✓ MCP server "claudedash" added
# Claude can now query its own dashboard
You → What tasks are left in the queue?
tool: get_queue
→ 8 READY · 7 DONE · 1 BLOCKED
Claude → 8 tasks are ready to start. S4-T3 is blocked on S4-T2. Would you like me to start S2-T4?
MCP Server

Claude can query
its own dashboard.

Add claudedash as an MCP server and your agent gains full self-awareness: what tasks are queued, how much context is left, what it's spent today. The loop closes.

get_queuetask list with READY/BLOCKED status
get_sessionsactive sessions + context health
get_costtoday's spend by model
log_taskwrite DONE/BLOCKED to execution.log
create_taskadd new task to queue.md

Start watching your agents.

One command. No config. Live Kanban, Agent API, quality analytics — in real time.

GitHub ↗