The infrastructure layer for AI agent swarms.
88 MCP tools · A2A protocol · OmniMesh VPN · nftables firewall · CDP browser · cookie sync · 2FA TOTP · bi-directional sync · CyberBase persistence
Quick Start
npm install -g omniwire
Add to your AI agent (Claude Code, Cursor, OpenCode, etc.):
{
"mcpServers": {
"omniwire": { "command": "omniwire", "args": ["--stdio"] }
}
}
Why OmniWire?
Use Cases
DevOps & Infrastructure
# Deploy to all nodes in one call
omniwire_deploy(src="node-a:/app/v2.tar.gz", dst="/opt/app/")
# Rolling service restart
omniwire_batch([
{node: "node1", command: "systemctl restart app"},
{node: "node2", command: "systemctl restart app"}
], parallel=false)
# Monitor disk across fleet
omniwire_disk_usage()
Security & Pentesting
# Anonymous nmap through Mullvad VPN
omniwire_exec(
node="node-a",
command="nmap -sV -T4 target.com",
via_vpn="mullvad:se",
background=true
)
# Rotate exit IP between scans
omniwire_vpn(action="rotate", node="node-a")
# Run nuclei through VPN namespace
omniwire_exec(command="nuclei -u target.com",
via_vpn="mullvad", store_as="nuclei_results")
Multi-Agent Coordination
# Agent A dispatches work
omniwire_task_queue(action="enqueue",
queue="recon", task="subfinder -d target.com")
# Agent B picks it up
omniwire_task_queue(action="dequeue", queue="recon")
# Share findings on blackboard
omniwire_blackboard(action="post",
topic="subdomains", data="api.target.com")
# A2A messaging between agents
omniwire_a2a_message(action="send",
channel="results", message="scan complete")
Background & Async Workflows
# Long build in background
omniwire_exec(
command="docker build -t app .",
node="node-a", background=true
)
# Returns: "BACKGROUND bg-abc123"
# Check progress
omniwire_bg(action="poll", task_id="bg-abc123")
# Returns: "RUNNING (45.2s)"
# Get result when done
omniwire_bg(action="result", task_id="bg-abc123")
# Pipeline: build → test → deploy
omniwire_pipeline(steps=[
{node: "node-a", command: "make build"},
{node: "node-a", command: "make test"},
{command: "deploy.sh", store_as: "version"}
])
File Operations
# Transfer large dataset between nodes
omniwire_transfer_file(
src="node-a:/data/model.bin",
dst="hostinger:/ml/model.bin"
)
# Auto-selects: aria2c (16-conn parallel)
# Sync config to all nodes
omniwire_deploy(
src_node="node-a",
src_path="/etc/nginx/nginx.conf",
dst_path="/etc/nginx/nginx.conf"
)
VPN & Anonymous Operations
# Full Mullvad setup for a node
omniwire_vpn(action="connect", server="se",
node="node-a")
omniwire_vpn(action="quantum", config="on")
omniwire_vpn(action="daita", config="on")
omniwire_vpn(action="multihop", config="se:us")
omniwire_vpn(action="dns", config="adblock")
omniwire_vpn(action="killswitch", config="on")
# Verify anonymous IP
omniwire_vpn(action="ip", node="node-a")
# Node-wide VPN (mesh stays connected)
omniwire_vpn(action="full-on", server="de")
Architecture
graph TB
subgraph clients["AI Agents"]
CC["Claude Code"]
OC["OpenCode / OpenClaw"]
CU["Cursor / Any MCP Client"]
A2["Other Agents (A2A)"]
end
subgraph omniwire["OmniWire MCP Server"]
direction TB
MCP["MCP Protocol Layer<br/>stdio | SSE | REST"]
subgraph tools["88 Tools"]
direction LR
EXEC["Execution<br/>exec run batch<br/>broadcast pipeline bg"]
AGENT["Agentic<br/>store watch task<br/>a2a events locks"]
FILES["Files & Deploy<br/>read write transfer<br/>deploy find"]
SYS["System & DevOps<br/>docker services<br/>cron env git syslog"]
SYNC["CyberSync<br/>sync diff search<br/>secrets knowledge"]
end
subgraph engine["Core Engine"]
direction LR
POOL["SSH2 Pool<br/>persistent compressed<br/>circuit breaker"]
XFER["Transfer Engine<br/>SFTP netcat+LZ4<br/>aria2c 16-conn"]
CSYNC["Sync Engine<br/>PostgreSQL XChaCha20<br/>parallel reconcile"]
end
end
subgraph mesh["Infrastructure Mesh"]
direction LR
N1["Node A<br/>storage"]
N2["Node B<br/>compute"]
N3["Node C<br/>GPU"]
N4["Node D<br/>local"]
end
DB[("PostgreSQL<br/>CyberBase")]
CC & OC & CU & A2 -->|MCP| MCP
MCP --> tools
tools --> engine
POOL -->|"SSH2 multi-path"| N1 & N2 & N3
POOL -->|"local exec"| N4
CSYNC --> DB
style omniwire fill:#0D1117,stroke:#59C2FF,stroke-width:2px,color:#C6D0E1
style clients fill:#161B22,stroke:#91B362,stroke-width:1px,color:#C6D0E1
style mesh fill:#161B22,stroke:#E6B450,stroke-width:1px,color:#C6D0E1
style tools fill:#0D1117,stroke:#59C2FF,stroke-width:1px,color:#C6D0E1
style engine fill:#0D1117,stroke:#CC93E6,stroke-width:1px,color:#C6D0E1
style MCP fill:#162B44,stroke:#59C2FF,color:#59C2FF
style DB fill:#162B44,stroke:#CC93E6,color:#CC93E6
Agent Setup Instructions
OpenClaw
OmniWire is available as a ClawHub skill:
# Install via ClawHub CLI
clawhub install omniwire
# Or manual: copy integrations/openclaw/SKILL.md to your OpenClaw skills directory
cp integrations/openclaw/SKILL.md ~/.openclaw/skills/omniwire.md
CyberSync automatically ingests OpenClaw agents, skills, memory, and workspace into CyberBase PostgreSQL.
PaperClip
Register OmniWire as a zero-cost infrastructure agent:
{
"agents": [{
"name": "omniwire",
"type": "local-cli",
"command": "omniwire --stdio",
"skills": ["mesh-exec", "file-transfer", "service-control", "docker", "vpn", "scraping", "firewall"],
"budget": { "monthly_usd": 0 }
}]
}
See integrations/paperclip/ for the full adapter and skill definition.
Claude Code
{
"mcpServers": {
"omniwire": {
"command": "omniwire",
"args": ["--stdio"]
}
}
}
OpenCode / Oh-My-OpenAgent
{
"mcp": {
"omniwire": {
"type": "local",
"command": ["omniwire", "--stdio"]
}
}
}
Codex / Gemini
CyberSync automatically syncs OmniWire config to Codex and Gemini environments.
Key Capabilities
Execution
omniwire_exec single command + retry + assert
omniwire_run multi-line script (compact UI)
omniwire_batch N commands, 1 tool call, chaining
omniwire_broadcast parallel across all nodes
omniwire_pipeline multi-step DAG with data flow
omniwire_bg poll/list background tasks
Multi-Agent (A2A)
omniwire_store session key-value store
omniwire_a2a_message agent-to-agent queues
omniwire_event pub/sub event bus
omniwire_semaphore distributed locking
omniwire_agent_task async background dispatch
omniwire_workflow reusable named DAGs
Adaptive File Transfer
< 10 MB SFTP native, 80ms
10M-1GB netcat+LZ4 compressed, 100ms
> 1 GB aria2c 16-parallel, max speed
Connection Resilience
Connected --> Health Ping (30s, parallel)
|
Failure --> Multi-path Failover
| WireGuard -> Tailscale -> Public IP
|
+--> Retry (300ms -> 600ms -> ... -> 10s)
|
3 fails --> Circuit OPEN (15s) -> Auto-recover
Background Dispatch
# Any tool supports background: true
exec(background=true) -> "bg-abc123"
bg(action="poll", id=..) -> "RUNNING (3.2s)"
bg(action="result", id=..) -> full output
bg(action="list") -> all tasks + status
Agentic Chaining
exec(store_as="ip") store result
exec(command="ping {{ip}}") interpolate
batch(abort_on_fail=true) fail-fast
exec(format="json") structured output
exec(retry=3, assert="ok") resilient
watch(assert="ready") poll until
All 88 Tools
Every tool supports
background: true— returns a task ID immediately. Poll withomniwire_bg.
Performance
- Cipher: AES-128-GCM (AES-NI hardware accelerated)
- Key exchange: curve25519-sha256 (fastest modern KEX)
- Keepalive: 2s interval, 2 retries = 4s dead detection
- Port finder:
shuf(pure bash) replacespython3 -c socket(-30ms) - Compression: LZ4-1 for transfers (10x faster than gzip)
- Buffer: Array push + join (O(n) vs O(n^2) string concat)
- Status: Single
/procread replaces multiple piped commands - Health ping:
:builtin (no hash lookup, no fork) - Reads: SFTP tried first,
catfallback only on failure - Circuit breaker: 15s recovery, 10s reconnect cap
Security
- All remote execution via
ssh2.Client.exec()-- neverchild_process.exec() - Key-based auth only, no passwords stored, SSH key caching
- Multi-path failover: WireGuard -> Tailscale -> Public IP
- Secret-bearing tool output is redacted by default (
omniwire_envlists keys, not values) - Prefer 1Password or another managed vault for production secrets; file/env backends are local-development fallbacks
- 2MB output guard prevents memory exhaustion
- 4KB auto-truncation prevents context window bloat
- Circuit breaker isolates failing nodes
- CORS restricted to localhost on REST API
Transport Modes
omniwire --stdio # MCP mode (default)
omniwire --sse-port=3200 --rest-port=3201 # HTTP mode
omniwire --stdio --no-sync # MCP without CyberSync
omniwire # or: ow # Interactive REPL
Configure Mesh
Create ~/.omniwire/mesh.json:
{
"nodes": [
{ "id": "server1", "host": "10.0.0.1", "user": "root", "identityFile": "id_ed25519", "role": "storage" },
{ "id": "server2", "host": "10.0.0.2", "user": "root", "identityFile": "id_ed25519", "role": "compute" }
]
}
Agentic Installation / Setup
For adding a new node to your OmniWire mesh — what to have ready, how to wire it in, and how to connect it to Claude Code.
Prerequisites
Install OmniWire
npm install -g omniwire
omniwire --version # verify
Add a New Node
1. Generate WireGuard keypair on the new node:
wg genkey | tee /etc/wireguard/node_private.key | wg pubkey > /etc/wireguard/node_pub.key
cat /etc/wireguard/node_pub.key
2. Assign it a mesh IP (next available in 10.0.0.0/24):
3. Register the node with OmniMesh (run from any node already in the mesh):
omniwire_omnimesh(action="add-peer",
id="newnode",
public_key="<pubkey from step 1>",
allowed_ips="10.0.0.N/32",
endpoint="<public IP or DNS>:51820"
)
4. Push updated peer list to all nodes:
omniwire_omnimesh(action="sync-peers")
5. Bring the interface up on the new node:
wg-quick up wg0
ping 10.0.0.1 # verify hub reachability
Configure OmniWire on the New Node
Add the node to ~/.omniwire/mesh.json (create if absent):
{
"nodes": [
{ "id": "node-a", "host": "10.0.0.1", "user": "root", "identityFile": "~/.ssh/id_omniwire", "role": "storage" },
{ "id": "hostinger", "host": "10.0.0.2", "user": "root", "identityFile": "~/.ssh/id_omniwire", "role": "compute" },
{ "id": "windows", "host": "10.0.0.3", "user": "Admin", "identityFile": "~/.ssh/id_omniwire", "role": "local" },
{ "id": "thinkpad", "host": "10.0.0.4", "user": "user", "identityFile": "~/.ssh/id_omniwire", "role": "local" }
]
}
Verify connectivity:
omniwire_mesh_status # should show all nodes green
omniwire_doctor # checks SSH, disk, mem, WireGuard, CyberBase
Connect Claude Code via MCP
Add to ~/.claude/claude_desktop_config.json (or your IDE's MCP config):
{
"mcpServers": {
"omniwire": { "command": "omniwire", "args": ["--stdio"] }
}
}
Restart Claude Code. Verify in a new session:
omniwire_mesh_status() # 88 tools should be available
Environment Variables
Set persistently on a node. Values are written to /etc/environment, but tool responses only confirm whether a value is set and never echo the secret back:
omniwire_env(action="set", key="OP_SERVICE_ACCOUNT_TOKEN", value="<token>", node="node-a")
omniwire_env(action="get", key="OP_SERVICE_ACCOUNT_TOKEN", node="node-a") # OP_SERVICE_ACCOUNT_TOKEN=<redacted: set>
CyberSync Auto-Distribution
CyberSync pushes configs, secrets, and memories to all nodes automatically via PostgreSQL.
# Check what's tracked
omniwire_coc(action="cybersync-status")
# Force push current config to all nodes
omniwire_coc(action="force-sync")
# Diff local state vs database
cybersync_diff()
On first run, CyberSync pulls node configs, 2FA seeds, and Claude memories from CyberBase — no manual copy-paste between machines.
Changelog
19 new tools: proxy, dns, backup, container, cert, user, schedule, alert, log_aggregate, benchmark, snippet, alias, trace, doctor, metrics, audit, plugin, cookies, cdp.
CyberBase auto-persistence: Store, audit, blackboard, cookies all sync to PostgreSQL. pgvector semantic search. 5s statement_timeout on all DB calls.
Architecture: Priority command queues, smart output truncation, predictive node selection, latency history, connection pool stats.
Security: Command denylist (blocks rm -rf /, fork bombs, disk wipes). Audit log with CyberBase persistence.
A2A: Typed message schemas (JSON validation), dead letter queue for failed tasks, pub/sub event filters.
DX: GitHub Actions CI, bash/zsh/fish shell completions, --json flag, cookie sync to 1Password.
omniwire_firewall: nftables-based firewall engine with 17 actions. Presets (server, paranoid, minimal, pentest), rate-limiting, geo-blocking by country, port-knocking sequences, IP ban/unban, whitelist/blacklist, rule management, audit log, save/restore.
Zero mesh impact: wg0, wg1, tailscale0, and all mesh CIDRs (10.0.0.0/24, 10.20.0.0/24, 100.64.0.0/10) are always whitelisted before any hardening rules. nftables runs in kernel space — zero latency overhead.
omniwire_vpn tool: Mullvad, OpenVPN, WireGuard, Tailscale. Split-tunnel (per-command) + full-node modes. Mesh connectivity (wg0, wg1, Tailscale) always preserved via route exclusions and network namespace isolation.
via_vpn on exec: Route any command through VPN using Linux network namespaces. Only the command's traffic goes through VPN — SSH/WireGuard mesh stays on real interface.
Modes: connect (split-tunnel), full-on (node-wide with mesh exclusions), rotate (new exit IP), status, list, ip.
background: true auto-injected into all 88 tools via server-level wrapper. Returns task ID, poll with omniwire_bg. New omniwire_bg tool for list/poll/result.
Performance: AES-128-GCM cipher, curve25519-sha256 KEX, 2s keepalive, LZ4 transfers (10x faster), shuf port finder (-30ms), SFTP-first reads, array buffer concat, /proc single-read status, : builtin health ping, 300ms reconnect start, 15s circuit breaker.
4 new A2A tools (49 -> 53): agent_registry (capability discovery), blackboard (swarm collaboration), task_queue (distributed work), capability (node routing).
9 new agentic tools (40 -> 49): store, pipeline, watch, healthcheck, agent_task, a2a_message, semaphore, event, workflow. Agentic upgrades: format:"json", retry, assert, store_as, {{key}} interpolation.
Output overhaul (auto-truncation, smart time, tabular multi-node). 6 new DevOps tools (cron, env, network, clipboard, git, syslog).
Security fixes, multi-path SSH failover, CyberBase integration, VaultBridge Obsidian mirror.
omniwire/
src/
mcp/ MCP server (88 tools, 3 transports)
nodes/ SSH2 pool, transfer engine, PTY, tunnels
sync/ CyberSync + CyberBase (PostgreSQL, Obsidian, encryption)
protocol/ Mesh config, types, path parsing
commands/ Interactive REPL
ui/ Terminal formatting
Requirements: Node.js >= 20 • SSH key access to nodes • PostgreSQL (CyberSync only) • WireGuard recommended