Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling

Anthropic's Claude Code integration with tmux — the guide demonstrates a “session‑per‑project” naming scheme, pane layouts, send‑keys scripting, capture‑pane reads, and a status‑line monitor, while also noting the practical ceiling of four concurrent Claude agents per tmux server—highlights how developers are now orchestrating multiple AI agents locally with precise control. For website owners this means AI‑driven bots are increasingly being managed as lightweight, scriptable agents that can crawl, scrape, or interact with sites from a single host. The four‑agent limit pushes developers to spin up additional tmux servers or container instances, potentially multiplying the number of distinct IPs and user‑agents hitting a site. As these agents mimic human browsing patterns (e.g., sending keystrokes, capturing output), traditional bot‑detection heuristics may miss them, raising the risk of undetected traffic spikes, content scraping, or automated form abuse. **Actionable tips:** 1. **Tag tmux‑spawned agents in llms.txt** – add a custom `User‑Agent: ClaudeAgent/​<session‑id>` entry for each session so your llms.txt file can explicitly allow or block these agents. 2. **Monitor tmux status lines via your analytics** – configure your log parser to capture the tmux status‑line strings (which include session names) and flag any traffic that matches the “project‑X‑agent‑Y” pattern. 3. **Rate‑limit per tmux server** – enforce a per‑IP or per‑User‑Agent request cap that aligns with the four‑agent ceiling, preventing a single host from exceeding expected load and triggering automated abuse alerts.