A tmux-backed agent pool with MCP server for orchestrating interactive CLI agents
Agent-mux provides 12 tools with basic schemas and descriptions, but falls short of production quality. All tools have descriptions (10-150 chars, mostly adequate) and input schemas with type definitions. However, parameter descriptions are sparse or missing entirely, most params lack guidance on format, constraints, or valid values. Output schemas are undocumented; callers cannot predict response structure. Error handling is minimal: no recovery guidance, no categorization of retryable vs. fatal errors. Tool composition is reasonable (each does one thing), but naming could be clearer (e.g., 'send_to_window' is vague vs. 'send_keys_to_tmux_window'). Security concerns: no mention of permission gates, audit logging, or input sanitization for tmux command injection. Baseline: 549 production tools average 194-char descriptions and 100% param coverage; agent-mux averages ~80 chars with ~40% param descriptions.
Switch to agent window (best-effort: switch tmux clients to project + select-window)
Close entire project (kill tmux session; keeps .ccx/projects/<project>/ files)
Create a new project (tmux session) and initialize .ccx/projects/<project>/ directory
Interrupt agent (Ctrl+C)
Close agent (kill tmux window + update registry)
List agents (defaults to current project; can specify project)
List all projects (tmux sessions + local registry directories)
Output schemas undocumented. No tool describes what fields it returns or their types. LLMs cannot predict response structure or plan downstream calls.
Parameter descriptions missing or incomplete. 'text' param in send_to_agent lacks format guidance. 'lines' in read_agent_output has no min/max bounds. 'options' in spawn_agent is an object with nested properties but no guidance on which are required or how they interact.
No error handling guidance. Tools throw generic errors ('Agent not found', 'tmux session not found') without recovery hints. LLMs cannot self-correct or know whether to retry.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
Read agent output (tail N lines from pipe-pane log file)
Send message to agent (via tmux send-keys)
Send message to any tmux window (not limited to registered agents, can send to main window)
Create agent (type: claude/codex/gemini), one agent = one tmux window
Switch current project (writes to .ccx/projects/current.json and best-effort switches tmux clients)
Destructive operations lack confirmation or dry-run. close_project and kill_agent irreversibly destroy state with no undo or confirmation step. Agents can accidentally wipe projects.
Command injection risk. send_to_agent and send_to_window accept arbitrary text and pass it to tmux send-keys. No sanitization visible; malicious input could break tmux syntax or execute unintended commands.