Skip to content
Back to skills

Swe Mux

ASecurity

Observe and coordinate with the swe-mux agent fleet this session runs inside: sibling sessions, live status, transcripts, history search, Project notes, bounded messages, settle waits. Use only when the user names swe-mux or mux, or asks about sibling sessions, the fleet, or another agent's work. Do not use merely because a task could benefit from delegation or a background terminal. Requires MUX_SESSION_ID.

  • 6 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 24, 2026
ai-agentsrustshellbash

Works with

  • terminal
  • cli
  • mcp

Security analysis

A100/100

Scanned September 24, 2026

npx -y skills add jatoran/swe-mux --skill swe-mux --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Swe Mux?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Swe Mux
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/jatoran-swe-mux/badge)](https://www.skillsdirectory.com/skills/jatoran-swe-mux)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: swe-mux
description: "Observe and coordinate with the swe-mux agent fleet this session runs inside: sibling sessions, live status, transcripts, history search, Project notes, bounded messages, settle waits. Use only when the user names swe-mux or mux, or asks about sibling sessions, the fleet, or another agent's work. Do not use merely because a task could benefit from delegation or a background terminal. Requires MUX_SESSION_ID."
metadata:
  managed-by: swe-mux
---

# swe-mux

swe-mux runs coding agents in real terminals and gives every session a bounded view of its
fleet: sibling sessions and their live status, transcripts, indexed history search, Project
notes, message queues a human can supervise, and bounded acts - messaging a sibling, waiting
for one to settle, requesting a spawn or a land.

Before acting on anything here, verify this agent is inside a swe-mux pane:

```bash
test -n "${MUX_SESSION_ID:-}"
```

If that check fails, say you are not running inside swe-mux and stop. Do not inspect or
drive swe-mux sessions from outside one.

## Which surface this session holds

The same capabilities travel on two transports, and `MUX_SURFACES` in this environment says
which this session has: `mcp,cli`, `mcp`, `cli`, or empty for neither (an operator switched
fleet access off for this harness - respect that and stop). An unset variable means an
older daemon; treat it as `mcp,cli`.

- **Prefer the `mux` MCP tools when they are in your tool list.** They carry this session's
  identity implicitly, need no shell quoting, and their descriptions are the current
  contract - trust them over anything remembered or written here.
- **Otherwise, with `cli` in `MUX_SURFACES`, use the CLI's agent mode.** `swemux agent
  tools` lists exactly the tools this session may call - the same tools, same authority
  checks, same budgets as MCP - and `swemux agent call <tool> key=value ...` calls one
  (`key:=value` for JSON-typed values, `--input` for a whole JSON object). A typed refusal
  prints the tool's own JSON and exits 1. Run `swemux --help` for the command surface, and
  read session ids, Project names, and message ids out of returned JSON rather than
  predicting them; treat an empty result as "nothing relevant exists in that scope", not as
  an error.

Two states that mislead: `idle` alone does not mean a session finished (a session awaiting
input and one idle behind background work render identically), and a message you stage is
delivered later under the receiving session's own rules, not when your call returns. To be
told when a sibling stops working, arm a watch; to block briefly on it, use the await tool -
its timeout is a normal, re-callable result, not an error.

## Operator commands are not yours

Every other `swemux` command acts as the human operator - sending raw input, killing or
spawning sessions, reloading or updating the daemon - and bypasses the queues, approvals,
and provenance that make agent-to-agent actions reviewable. The daemon refuses the
session-acting ones from an agent pane and names the agent surface instead. Do not run them
unless the human operator explicitly asked for exactly that.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…