Use case / AI agents

Let AI tools call approved services, not inherit their secrets.

Route a request to a named agent on a specific device. The destination authenticates the caller, checks the action, and uses its own local service credential.

For developers connecting coding assistants and internal automation to selected MCP servers, APIs, or databases.

Explore the agent guide

Quick answers

Can a coding assistant call a private MCP tool?

Yes, after Airmsh and the coding-assistant integration are configured, the destination agent exists, and the caller is allowed for the requested tool and action. Installing a plugin alone does not grant access. See agent setup.

How is the destination agent identified?

By both target device ID and agent tag. A network Endpoint is only a path, never proof of agent identity. The destination verifies the source device, route, action, and resource. See agent identity and pinning.

Does the remote assistant receive database or MCP credentials?

No. It sends a request; the destination device resolves its own secret reference and authenticates locally. Native service permissions still apply. See the security model.

Device plus agent tag

An agent route is identified by a target device ID and agent tag together. A network Endpoint is only the transport path; it is not the agent identity. The receiving gateway checks the authenticated source device, target, tag, protocol, action, and resource.

Remote callers send the query, prompt, or tool call. Database passwords and MCP tokens are resolved on the destination device for native or local authentication, not returned to the caller.

Policy before execution

The destination applies route and capability checks, request limits, local-service authentication, sensitive-data decisions, response limits, and signed audit metadata in that order. An owner can enable additional per-agent policy and pin an MCP tool schema; an unexpected pinned schema change is denied until reviewed.

Illustrative workflow

What this looks like

A coding assistant needs the tools/list result from an internal MCP service. An owner registers a destination-local MCP agent, allows the specific caller, and optionally pins the approved tool schema. The assistant requests the agent by device ID and tag. A changed schema or disallowed tool call is rejected before execution.

Limits to know

Installing a Codex or Claude Code plugin does not itself grant access. Airmsh must be installed and signed in, and the destination agent and policy must be configured.
A native database or MCP service may impose its own authentication and permissions. Airmsh does not replace those controls.
Policy behavior depends on the deployed client, daemon, and server versions; verify the intended workflow with a test agent before using sensitive data.

Next step

Try the workflow on your own devices.