Claude Esp

claude-esp

Stream Claude Code's hidden output (thinking, tool calls, subagents) to a separate terminal in real-time.

Looking for the Rust port? See: claude-esp-rs

The Problem

When using Claude Code interactively, tool outputs and thinking are collapsed by default and require pressing Ctrl+O to toggle visibility. This tool lets you watch all that output in a separate terminal with a nice TUI, without interrupting your main Claude Code session.

Features

- Multi-session support - Watch all active Claude sessions simultaneously
- Hierarchical tree view - Sessions with nested Main/Agent nodes
- Real-time streaming - See thinking, tool calls, and outputs as they happen
- Subagent tracking - Automatically discovers and displays subagent activity
- Session events - Compaction boundaries, hook output (including hook errors, blocked stops and injected context), post-edit LSP diagnostics, PR-link and artifact-link events, session-cost summaries (cost, lines changed, models used), session continuations, background-task progress, prompt-cache misses (with reason + extra tokens), queue operations, slash-command invocations, transient notices, killed-subagent markers, model refusals, scheduled-task (cron / loop) fires, plan/auto mode transitions, permission-mode changes, and skill/MCP/tool deltas surfaced inline
- Agent type labels - Shows agent types (Explore, code-reviewer, etc.) from .meta.json
- Token usage tracking - Cumulative input/output token counts in the header bar
- Per-agent context size - Each Main/subagent row shows current context as a percentage of the model's max context window (Main 18%, Explore 9%). Denominator is the model's *max window* (1M for the Claude 5 family, Haiku 5.5 included, and opus-4-6 through 4-8 / sonnet-4-6, 200k for haiku-4-5), not the…