zswarm-mcp
Enables coordination of CLI AI agent crews across Zellij terminal panes, allowing users to spawn agents, send prompts, inspect outputs, sync signals, and manage git worktrees via natural language.
README
zSwarm
Coordinate CLI AI crews across Zellij terminal panes. Control agents, pass prompts between panes, manage git worktrees, inspect outputs, and sync crew tasks via CLI or Model Context Protocol (MCP).
you ──▶ any agent ──┐
│ zswarm({op:"send", to:"reviewer"})
▼
every pane can call zswarm ──▶ ┌─────────┐
so agents drive each other │ zswarm │
│ cli/mcp │
└────┬────┘
paste + Enter ┌────┴────┐ read screen / pushed state
▼ ▼
══ this machine ══════════════════════════════╗ ══ ssh user@build-01 ═════════╗
zellij "crew" ║ zellij "ci" ║
║ ║
┌────────────┐ send ┌────────────┐ ║ ┌────────────┐ ║
│ codex │────────▶│ claude │───────╫──▶│ opencode │ ║
│ builder │◀────────│ reviewer │ send ║ │ test-runner│ ║
│ wt:feat-a │ signal │ wt:rev-a │ ║ │ wt:feat-a │ ║
└────────────┘ └────────────┘ ║ └──────┬─────┘ ║
┌────────────┐ ┌────────────┐ ║ ┌──────▼─────┐ ║
│ cursor │ │ agy │ ║ │ pi │ ║
│ refactor │ │ docs │ ║ │ deploy │ ║
│ wt:feat-b │ │ wt:docs │ ║ │ cwd:/srv │ ║
└────────────┘ └────────────┘ ║ └────────────┘ ║
═══════════════════════════════════════════════╝ ══════════════════════════════╝
one pane = one agent + one role + optionally its own worktree
any pane can drive any other, on this machine or across the ssh boundary
⚡ Quick Start
Requirements
- Zellij ≥ 0.42 on
PATH(or setZSWARM_BIN/ZSWARM_PATH) - Node.js ≥ 20
Installation
npm i -g zswarm
zswarm list
Enable the optional Zellij event-bus plugin once per machine for zero-polling state updates:
zswarm bus --install
MCP Server Setup
Configure zswarm-mcp (stdio MCP server bundled with the package) in your MCP host:
{
"mcpServers": {
"zswarm": {
"command": "/absolute/path/to/zswarm-mcp",
"env": {
"ZSWARM_SESSION": "my-zellij-session"
}
}
}
}
Note: Use the absolute path to
zswarm-mcpand specifyZSWARM_SESSIONto prevent ambiguous session errors when multiple Zellij sessions run.
Usage & Quick Syntax
Every operation is supported both via CLI (zswarm <op> [flags]) and MCP (zswarm({ op: "<op>", ... })).
Essential Examples
# 1. Inspect Panes & Status
zswarm list # List active panes
zswarm status # Check status: busy | waiting | idle | exited
# 2. Spawn Panes & Worktrees
zswarm spawn --command "claude" --name reviewer # Spawn agent pane
zswarm spawn --command "codex" --worktree feature-auth --name auth-dev # Isolated git worktree pane
# 3. Pass Prompts & Input
zswarm send --to reviewer --body "Review changes in src/" --submit auto
zswarm keys --to reviewer --key "Ctrl c" # Send special keys / interrupts
# 4. Read Outputs & Wait
zswarm dump --to reviewer --max 4000 # Read recent screen tail
zswarm tail --to reviewer # Incremental read since last check
zswarm wait --to reviewer --match "DONE" # Block until pattern matched or pane idle
# 5. Coordinate Crew & Sync Signals
zswarm broadcast --all --group builder --body "Run test suite" # group matches pane title or command
zswarm signal --channel tests --payload ok # Send signal to durable channel
zswarm await --channel tests --count 3 # Wait for 3 worker signals
📋 Operations Reference
| Operation | Purpose | Key Parameters / Flags |
|---|---|---|
list / sessions |
List panes or live Zellij sessions | verbose |
status |
Panes classified by state (idle, busy, waiting, exited) |
to, sampleMs, since-last |
spawn |
Launch command in new pane, tab, or git worktree | command, name, cwd, worktree, tab, newTab |
send |
Send text / prompt into a pane | to, body, submit (auto|double-enter|none), expect |
dump |
Read screen content (screen tail) | to, max (default 8000) |
tail |
Incremental screen read since last cursor | to, reset |
wait |
Block until quiet, pattern match, or idle | to, match, for (idle), idleMs, timeoutMs |
broadcast |
Send one prompt to multiple panes | to, tab, all, group, body |
keys / interrupt |
Send key combos (Ctrl c, Esc) or raw chars |
to, keys, chars, enter, hard |
close |
Close a pane | to, force |
signal / signals / await |
Durable message channels & barrier sync | channel, payload, count |
worktrees / unworktree |
List or remove managed git worktrees | cwd, branch, path, force |
diff / checkpoint |
Inspect peer worktree patch or autocommit WIP | branch, path, stat, message |
rename / focus |
Retitle pane/tab or focus pane | to, tab, name |
tabs / layout / stack |
Inspect tabs, layout KDL, or stack panes | to, max |
bus / log |
Event bus plugin management & delivery log | install, clear, failed, limit |
🛡️ Safety & Policy Guards
- Self-Target Guard: Operations refuse to modify zSwarm's own pane (
allowSelf: trueto override). - Execution & Exited Guards: Prevents writing to exited panes (
force: trueto override). - Prompt Safety (
expect):send --expect <text>ensures target pane screen contains<text>before writing (prevents accidental shell execution). - Submission Verification (
submit):autoverifies text submission into TUI composers and retries Enter if unsubmitted.
Environment Security Controls
| Env Variable | Effect |
|---|---|
ZSWARM_READONLY=1 |
Disable all write ops (send, spawn, close, etc.) |
ZSWARM_ALLOW_PANES / ZSWARM_DENY_PANES |
Comma-separated list/patterns of allowed/denied pane titles |
ZSWARM_ALLOW_SPAWN=0 |
Disable spawning new panes |
ZSWARM_ALLOW_CLOSE=0 |
Disable closing panes |
ZSWARM_ALLOW_WORKTREE_REMOVE=0 |
Disable removing git worktrees |
⚙️ Environment Variables
| Variable | Description |
|---|---|
ZSWARM_BIN / ZSWARM_PATH |
Path to zellij executable |
ZSWARM_SESSION |
Default Zellij session name |
ZELLIJ_SESSION_NAME |
Active Zellij session (set automatically inside Zellij panes) |
ZELLIJ_PANE_ID / ZSWARM_SELF_PANE |
Sending pane id (Zellij sets the first inside a pane). Used to refuse self-target and to default from to that pane's title |
ZSWARM_FROM |
Sender label when from is omitted (overrides the pane title) |
ZSWARM_WORKTREE_ROOT |
Directory for linked worktrees (default: <repo>-worktrees) |
ZSWARM_STATE_DIR |
State storage directory for logs, signals, and cursors (default: ~/.zswarm) |
ZSWARM_SSH |
user@host — Route Zellij commands across SSH |
ZSWARM_REMOTE_BIN |
Remote zellij binary (zellij.exe on Windows) |
ZSWARM_TMP |
Remote IPC temp directory, or auto to parse live zellij --server paths |
ZSWARM_SSH_MODE |
interactive — Windows only: run each CLI call in the desktop session (scheduled task, same idea as schtasks /IT) |
ZSWARM_REMOTE_SHELL |
cmd or sh — override remote quoting (inferred from .exe / Windows tmp / interactive) |
ZSWARM_SERVE |
127.0.0.1:9419 — Forward ops to zswarm serve (SSH tunnel to loopback; non-loopback listen is refused) |
ZSWARM_SERVE_TOKEN |
Shared secret for zswarm serve. Required even on loopback (another local OS user can connect to 127.0.0.1); send the same value on the client |
Remote crews
Linux/macOS SSH works when it is the same user and $TMPDIR. Windows OpenSSH is session 0; live Zellij is usually the interactive desktop session. ZSWARM_TMP=auto points at that TEMP (enough to list sessions). Pane attach on Windows uses named pipes in the desktop session, so use ZSWARM_SSH_MODE=interactive or run zswarm serve next to Zellij.
# Linux/macOS, same user:
ZSWARM_SSH=user@host zswarm list
# Point SSH at the interactive TEMP (auto reads zellij --server …)
ZSWARM_SSH=user@host ZSWARM_TMP=auto zswarm sessions
# Windows: run the CLI inside the desktop session. Discovers TEMP unless ZSWARM_TMP is set.
ZSWARM_SSH=user@host ZSWARM_SSH_MODE=interactive zswarm list
# Any OS: run zswarm next to Zellij; another machine talks over a tunnel to 127.0.0.1
# On the host, in the session that owns Zellij (loopback only; a token is not a bind substitute):
ZSWARM_SERVE_TOKEN=secret zswarm serve --listen 127.0.0.1:9419
# Windows, once (bakes ZSWARM_SERVE_TOKEN into the logon task env): zswarm serve --install
# On the client:
ssh -fN -L 9419:127.0.0.1:9419 user@host
ZSWARM_SERVE=127.0.0.1:9419 ZSWARM_SERVE_TOKEN=secret zswarm list
file: plugin URLs stay on the machine that owns Zellij. A client with ZSWARM_SERVE never sends a local wasm path across the tunnel.
{
"mcpServers": {
"zswarm": {
"command": "/absolute/path/to/zswarm-mcp",
"env": {
"ZSWARM_SERVE": "127.0.0.1:9419",
"ZSWARM_SERVE_TOKEN": "secret",
"ZSWARM_SESSION": "crew"
}
}
}
}
📦 Monorepo Structure
| Package | Purpose |
|---|---|
zswarm |
Main meta-package (npm i -g zswarm) |
@zswarm/core |
Core Zellij client and shared dispatch engine |
@zswarm/cli |
zswarm command line interface |
@zswarm/mcp |
Stdio MCP server (zswarm-mcp) |
@zswarm/wasm |
Precompiled Rust event-bus WASM plugin |
@zswarm/pi |
Extension bridge for harnesses without native MCP support |
📜 License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。