Skip to content

MCP.md

Latest commit

 

History

History
59 lines (45 loc) · 2.13 KB

MCP.md

File metadata and controls

59 lines (45 loc) · 2.13 KB
Raw

Orbis over MCP

Orbis speaks the Model Context Protocol on stdin/stdout, so an assistant can query the network and, when allowed, change it.

The tools are not a separate surface. They are the same catalogue the built-in assistant uses, executed through the same backend, because two independently maintained tool sets drift: the one used less gets stale descriptions, loses a parameter, and eventually lies about what it does.

Running it

orbisd -mcp -config /etc/orbis/orbis.yaml            # 10 read-only tools
orbisd -mcp -mcp-write -config /etc/orbis/orbis.yaml # 21 tools, 11 mutating

Write access is off by default. With it off the mutating tools are not offered at all rather than offered and refused, which is the same choice the built-in assistant makes: a model told it can do something and then blocked produces a worse conversation than one that was never given the capability.

Logs go to stderr. Nothing else may write to stdout, which carries the JSON-RPC stream.

Claude Code

{
  "mcpServers": {
    "orbis": {
      "command": "ssh",
      "args": ["root@192.168.50.221", "orbisd", "-mcp", "-config", "/etc/orbis/orbis.yaml"]
    }
  }
}

Run it over SSH rather than exposing a port: the tools read the whole DNS history and device inventory, and the node already trusts an SSH session.

Tools

Read: get_network_summary, list_clients, list_connections, top_destinations, dns_log, top_blocked_domains, list_events, list_firewall_rules, system_status, list_ad_candidates.

Write (with -mcp-write): block_domain, allow_domain, add_firewall_rule, set_rule_enabled, delete_firewall_rule, apply_firewall, set_client_blocked, label_client, decide_ad_candidate, flush_dns_cache, refresh_blocklists. Every one is written to the audit log with the actor mcp.

One caveat the server states in its own instructions

A node that is not on the traffic path records only its own traffic and broadcast noise. Flows will look sparse and every source will be the node itself. That is a placement problem, not a quiet network, and the difference matters before drawing any conclusion from the data.