How It Works

clnode uses Claude Code’s built-in hook system to intercept agent lifecycle events and route context through a local DuckDB database.

Architecture

Claude Code Session
│
├── Agent A starts  ──→  hook ──→  clnode daemon ──→  DuckDB (store)
├── Agent A stops   ──→  hook ──→  clnode daemon ──→  DuckDB (save summary)
├── Agent B starts  ──→  hook ──→  clnode daemon ──→  DuckDB (read A's summary)
│                                       │
│                                       └──→ stdout: additionalContext
│                                             (A's results injected into B)
└── Leader only sees final reports — context stays minimal

Hook Events

clnode intercepts these Claude Code lifecycle events:

Event	When	What clnode Does
SessionStart	Claude Code session begins	Register session, link to project
SubagentStart	Agent spawned	Register agent, inject smart context via additionalContext
SubagentStop	Agent finishes	Save agent’s work summary to DB
PostToolUse	Edit/Write tool used	Track file changes
UserPromptSubmit	User sends a message	Attach project context (active agents, tasks, decisions)

The Key Insight

Agents don’t talk to each other directly. They talk through time.

Agent A finishes and leaves a summary in DuckDB. Agent B starts later and receives that summary automatically through the hook system. The hook is the message bus, DuckDB is the mailbox.

Components

Component	Role
hook.sh	Shell script that reads stdin, POSTs to daemon, outputs to stdout
Daemon	Hono server on port 3100 — processes hook events, serves Web UI
DuckDB	Local database storing sessions, agents, context, tasks
Skills	LLM instructions preloaded into agents (compress-output, etc.)
Rules	Project-wide conventions auto-loaded every conversation