agents

A local dashboard for running multiple Claude Code agents in parallel — each with its own working directory, direction (inline prompt or a markdown file it re-reads each turn), and keep-alive supervisor that auto-restarts on crash and auto-resumes when usage limits reset.

Built to replace a launchd-based heartbeat with something visible and controllable from a UI.

Stack

• Next.js 16 (App Router, Turbopack) on port 4000 — UI, file I/O API routes
• Bun WebSocket + HTTP server on port 4001 — agent supervisor, transcript streaming, file-change events
• Claude Code CLI spawned per agent (--dangerously-skip-permissions, configurable model / fallback / effort)
• Tailwind v4

Features

• Multiple agents running concurrently, each with its own workingDir
• Two direction modes: inline prompt, or "re-read this markdown file each turn"
• Live transcript per agent (JSONL session file watching), with pause-on-scroll-up autoscroll
• Keep-alive supervisor: restarts on crash, parses usage limit reached|<ts> from stderr and auto-resumes after the reset
• Per-agent file activity + diff panel
• Pinned notes as tabs (fuzzy vault search when adding)
• Registry persisted to data/agents.json; PID files let the supervisor reattach to surviving children across reloads

Run

bun install
bun dev            # starts Next (4000) + ws-server (4001) under concurrently

Open http://localhost:4000.

The supervisor detaches children and unrefs them, so bun --watch reloads don't kill running agents — it re-attaches via PID files on next init.

Cost

Agents run against the Claude Code CLI's OAuth credentials (Max plan). The supervisor explicitly strips ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, and CLAUDE_CODE_OAUTH_TOKEN from the child env to make sure no leaked API key can route a spawned agent to pay-per-token billing.