CC Harness

CC Harness

See what Claude Code is actually doing

An open-source desktop workbench for Claude Code: configure · debug · observe · orchestrate

简体中文 · Report an issue

Claude Code has grown from a single-session tool into a complex system: multi-session orchestration, background loops, scheduled wakeups, auto-memory. Every new capability adds another layer of opacity. The CLI shows you a counter — Running agent 47/200. CC Harness shows you the topology.

What it does

- Live subagent call tree — tails session jsonl in real time and renders the 5-level subagent/workflow graph with per-node latency, token cost, and nesting depth. When a workflow stalls, you see which branch is stuck.

- Hook sandbox — dry-run any hook with simulated input: stdout, stderr, exit code, and the transformed result, without opening a real session. Fully isolated environment (PATH / HOME / TMPDIR only — no API tokens, no credentials).

- Token cost breakdown — per-session panel splitting skills / subagents / MCP / plugins / base session. Click a slice to rank that bucket's most expensive turns, then click a row to land on that exact message in the replay. Includes a real-time optimizer: repricing the current session's Opus tokens at Sonnet rates, so you see the exact dollar amount you'd save — from real data, not an estimate.

- Loop & background task monitor — aggregates ScheduleWakeup events across all sessions, classified as pending / fired / expired, with trigger history for each loop.

- Skill trigger analyzer — extracts trigger keywords from each skill (classified by action / technology / format / topic) and shows example prompts that would activate it. Plus a Mermaid structure diagram of every skill.

- Dependency graph — maps five relation types (Skills → MCP, Hooks → MCP, Skills ↔…