Tmux Agents

Manage background coding agents in tmux sessions. Spawn Claude Code or other agents, check progress, get results.

Security Scan

The skill mostly does what it says (wraps tmux sessions to run local or cloud agent CLIs) but contains a few unexplained or risky choices (hardcoded working directory, flags that auto-accept/auto-apply, and reliance on external CLIs/credentials not declared) that you should review before installing.

Purpose & Capability

Name/description, required binary (tmux), the brew install of tmux, and included scripts align with a tool that spawns and manages tmux-based agent sessions. Using external agent CLIs (claude, codex, gemini, ollama) is consistent with the stated purpose.

Instruction Scope

spawn.sh sends commands into tmux that launch external CLIs and runs them with flags like --dangerously-skip-permissions and --auto-edit/--full-auto which may cause agents to accept permissions or auto-apply edits without interactive confirmation. The script also unconditionally cds to ~/clawd (a hardcoded path) which is not documented as required and may cause unexpected behavior or operate on files in that directory.

Install Mechanism

Install uses a Homebrew formula for tmux, a well-known package source; nothing is downloaded from an arbitrary URL and no archive extraction is used. This is low risk.

Credentials

The skill declares no required env vars or credentials, but the runtime expects external agent CLIs that typically require API keys/config (cloud agents). Those credentials are not requested or documented as required environment variables — the scripts assume the user has preconfigured CLIs. Also the hardcoded ~/clawd path is an implicit config that wasn't declared.

Persistence & Privilege

always is false and the skill does not request elevated or persistent platform-wide privileges, nor does it modify other skills or global agent settings.

What to consider before installing

This skill is coherent with its stated purpose, but review and consider the following before installing or running it: - Review scripts (spawn.sh) locally: it cd's to ~/clawd — adjust or remove that if you don't want the agent working directory to be your home/project path. - Be cautious about the agent CLI flags: --dangerously-skip-permissions, --auto-edit and --full-auto can make agents accept permissions or apply edits automatically. Use them only if you trust the agent and understand the consequences. - The skill relies on external CLIs (claude, codex, gemini, ollama). Ensure those CLIs are installed and configured with the correct API keys/credentials; the skill does not declare or manage those credentials. - Treat agent names and custom commands as executable input: if you pass untrusted strings as the agent argument they will be invoked as commands inside tmux. Avoid using this skill with untrusted inputs. - Test in a safe environment first (a throwaway repo or container) and avoid running as a privileged user. If you want to limit risk, remove or alter the auto-approve flags and the hardcoded cd before using. If you want me to, I can suggest a safer spawn.sh variant that omits auto-approve flags, validates inputs, and uses a configurable working directory instead of ~/clawd.

Run coding agents in persistent tmux sessions. They work in the background while you do other things.

Available Agents

☁️ Cloud Agents (API credits)

Agent	Command	Best For
claude	Claude Code	Complex coding, refactoring, full projects
codex	OpenAI Codex	Quick edits, auto-approve mode
gemini	Google Gemini	Research, analysis, documentation

🦙 Local Agents (FREE via Ollama)

Agent	Command	Best For
ollama-claude	Claude Code + Ollama	Long experiments, heavy refactoring
ollama-codex	Codex + Ollama	Extended coding sessions

Local agents use your Mac's GPU — no API costs, great for experimentation!

Quick Commands

Spawn a new agent session

./skills/tmux-agents/scripts/spawn.sh <name> <task> [agent]

# Cloud (uses API credits)
./skills/tmux-agents/scripts/spawn.sh fix-bug "Fix login validation" claude
./skills/tmux-agents/scripts/spawn.sh refactor "Refactor the auth module" codex
./skills/tmux-agents/scripts/spawn.sh research "Research caching strategies" gemini

# Local (FREE - uses Ollama)
./skills/tmux-agents/scripts/spawn.sh experiment "Rewrite entire test suite" ollama-claude
./skills/tmux-agents/scripts/spawn.sh big-refactor "Refactor all services" ollama-codex

List running sessions

tmux list-sessions
# or
./skills/tmux-agents/scripts/status.sh

Check on a session

./skills/tmux-agents/scripts/check.sh session-name

Attach to watch live

tmux attach -t session-name
# Detach with: Ctrl+B, then D

Send additional instructions

tmux send-keys -t session-name "additional instruction here" Enter

Kill a session when done

tmux kill-session -t session-name

When to Use Local vs Cloud

Scenario	Recommendation
Quick fix, time-sensitive	☁️ Cloud (faster)
Expensive task, budget matters	🦙 Local
Long experiment, might fail	🦙 Local
Production code review	☁️ Cloud (smarter)
Learning/exploring	🦙 Local
Heavy refactoring	🦙 Local

Parallel Agents

Run multiple agents simultaneously:

# Mix and match cloud + local
./scripts/spawn.sh backend "Implement user API" claude           # Cloud
./scripts/spawn.sh frontend "Build login form" ollama-codex      # Local
./scripts/spawn.sh docs "Write API documentation" gemini         # Cloud
./scripts/spawn.sh tests "Write all unit tests" ollama-claude    # Local

Check all at once:

./skills/tmux-agents/scripts/status.sh

Ollama Setup

Local agents require Ollama with a coding model:

# Pull recommended model
ollama pull glm-4.7-flash

# Configure tools (one-time)
ollama launch claude --model glm-4.7-flash --config
ollama launch codex --model glm-4.7-flash --config

Tips

• Sessions persist even if Clawdbot restarts
• Use local agents for risky/experimental work
• Use cloud for production-critical tasks
• Check tmux ls to see all active work
• Kill sessions when done to free resources