Public MCP and external automation

Public MCP lets Codex, Claude Desktop, Claude Code, and compatible clients call the currently running Navop. It is not a fixed cloud API: discovery uses a dynamic loopback port and a user-only token, while live tools and schemas come from the local application.

Choose a server mode

Temporary mode is appropriate for short-lived use; clients should not expect it to remain available after the task or app ends. Persistent mode supports clients that need continuing local discovery. Both should remain loopback-only and use the protected per-user discovery information.

Do not expose discovery files, tokens, or ports to the public network. Changing mode, restarting Navop, or changing Tool Exposure may restart the endpoint, requiring MCP and ACP clients to reconnect.

Set permissions and Tool Exposure

Safe, Confirm, and Auto profiles control approval behavior. Begin with Safe or Confirm, and use Auto only when the task, target, and tools are tightly controlled. These profiles do not replace database or server authorization.

Tool Exposure independently controls Terminal, SSH Exec, visible terminal, Connections, SFTP, Redis, MongoDB, Database, and internal functions. Enable only the groups needed for the current client, then turn them off afterward.

Install client prerequisites

The bridge requires Node.js 20+ and working npx. Verify the version, then copy the generated configuration for Codex, Claude Desktop, Claude Code, or a generic MCP JSON client from Navop Settings. Follow each client's path and restart instructions.

Navop can also install or update its Skill for Codex and Agents. Install @navop/cli globally before using the Skill; an AI Agent then uses navop ... --json to discover and operate databases, SSH sessions, terminals, files, connections, and workspaces owned by Navop. The Skill does not place a static copy of every tool in the prompt; it teaches the Agent to inspect status, commands, and live schemas only when needed.

Why use the Navop Skill

When Navop is registered as a native MCP server, a client may send many enabled tool names, descriptions, and JSON schemas to the model on every turn. As the catalog grows, those repeated definitions consume context and Tokens. The Navop Skill gives an AI Agent a compact workflow and runs navop status, domain, or tool schema/call commands only when the task requires them.

npm install -g @navop/cli@latest
navop skill install --target codex --scope user
navop status --json
navop db query --help
navop tool schema <tool-name> --json
navop tool call <tool-name> --arguments '<json-object>' --json

Representative read-only operations are shown below. Replace placeholders with connection or session ids returned by the running Navop app, and inspect --help or the live schema before execution:

navop connections list --json
navop connections sessions --json
navop ssh exec --target <ssh-session-id> --command 'uname -a' --json
navop sftp list --connection <ssh-connection-id-or-name> --path /var/log --json
navop redis get --connection-id <redis-connection-id-or-name> --key app:status --json
navop mongo find --connection-id <mongo-session-id> --database app --collection users --filter '{"active":true}' --limit 20 --json
navop db query --connection <database-connection-id-or-name> --sql 'SELECT 1' --json
navop terminal read --target <terminal-session-id> --lines 80 --json

This is useful for terminal-capable Agents such as Codex: they can discover current Navop resources and operations on demand without pre-registering the complete Navop catalog in every model turn. It can reduce repeated context and token overhead, although the exact savings depend on how the client injects MCP tool definitions and how many groups are enabled.

The Skill does not completely remove MCP from the transport. The navop CLI still connects internally to Navop's authenticated loopback Public MCP endpoint, and Navop remains responsible for Tool Exposure, resource ids, permissions, approvals, sessions, results, and auditing. The Skill changes the Agent-side interaction from “carry every tool each turn” to “discover and call through terminal commands on demand.”

Use @navop/cli

@navop/cli provides status, tools, schema, call, and resource-oriented domain commands. The separate @navop/mcp package only runs the stdio bridge for compatible MCP clients.

Run navop ... only after checking the package source. Always obtain resource IDs and argument definitions from the current tools/schema output. Never guess IDs, reuse IDs from another device, or attempt to bypass approval.

Check or update the resolved CLI when needed:

npm view @navop/cli version
navop --version

Run npm install -g @navop/cli@latest before using the Skill, and npm update -g @navop/cli to update it. Native MCP clients should use npx -y @navop/mcp@latest as their stdio bridge.

Approve and troubleshoot

The approval window shows the requesting client, operation, resource, and parameters. Reject unexpected requests and correct the client instead of globally weakening permissions. ACP authorization does not automatically approve Public MCP.

For failures, check that Navop is running, server mode, Node.js, client configuration, discovery permissions, Tool Exposure, and live schema. Reconnect after endpoint changes. Redact tokens, local paths, connection names, and business parameters from logs.