AppEngine exposes a Model Context Protocol endpoint. Connect an HTTP MCP client to discover registered business services, inspect their contracts and call them in your organization. You configure the connection; the assistant handles tool discovery and calls.
Video tutorials for this topic
Endpoint
/mcp| Environment | URL | Status |
|---|---|---|
| Hosted AppEngine | https://appengine.appmint.io/mcp | Public tool discovery verified 11 October 2026 |
| Your own deployment | Your AppEngine base URL followed by /mcp | Verify against that deployment |
Choose Streamable HTTP in clients that ask for a transport. The endpoint is stateless: it accepts protocol messages over HTTP POST. It does not offer a persistent SSE stream; opening the URL in a browser with GET returns 405 and does not by itself indicate a broken MCP connection.
The protocol supports initialize, tools/list, tools/call and resources/list. The hosted discovery check above does not establish authenticated access to your organization's records.
Authentication
Set these connection headers for business-record calls:
orgid: <org>
Authorization: Bearer <token>Use a staff token from the intended organization. Keep it in the client's secret storage or an environment variable, not in an AI prompt. Staff authentication explains sign-in and its possible outcomes. Complete any required verification or password-change challenge; only a completed session supplies the access token.
Discovery (initialize, tools/list, list_services and describe_service) can work without credentials. call_service requires an authenticated principal, orgid and content read permission. A discovered tool therefore does not prove that record access is configured.
Current authorization boundary: MCP invokes registered services directly. Its gateway does not repeat every REST controller's menu or operation guard. Service-level checks still apply, but do not assume that a REST role restriction is automatically enforced through MCP. call_service can invoke read and write methods; a client approval prompt is not a server permission policy.
Client configuration
{
"mcpServers": {
"appengine": {
"url": "https://appengine.appmint.io/mcp",
"headers": {
"orgid": "<org>",
"Authorization": "Bearer <token>"
}
}
}
}Client configuration formats differ. The URL and headers above are the AppEngine connection values; use your client's supported HTTP MCP settings.
For Codex, the equivalent entry in ~/.codex/config.toml is:
[mcp_servers.appengine]
url = "https://appengine.appmint.io/mcp"
bearer_token_env_var = "APPENGINE_STAFF_TOKEN"
http_headers = { orgid = "YOUR_ORGANIZATION_ID" }Start the client with APPENGINE_STAFF_TOKEN available in its environment. Use your deployment's URL when it differs from the hosted address. codex mcp get appengine shows the configured entry; a successful tool invocation verifies the connection itself. After changing the entry or token environment, start a fresh client session so it uses those values. The official Codex MCP guide documents current supported settings.
The three tools
| Tool | What it does |
|---|---|
list_services | Every backend service exposed to MCP, with a one-line description and the methods each offers |
describe_service | Method signatures for one service — parameter names, types, which are required, and the return shape |
call_service | Invoke a method on a registered service |
The shape is deliberate: a client discovers what exists, reads the signature of the method it wants, then calls it. An agent does not need the API memorized or hard-coded, and new services become reachable without the client changing.
Arguments are passed in declaration order, with orgId omitted — the tenant comes from the header, not the argument list. Read the signature with describe_service first rather than guessing.
Checking it works
Open your AI client's MCP tool list and look for list_services, describe_service and call_service. Ask a small read-only question about a training record you may access, such as “Read this employee and tell me their name and staff code. Do not change anything.” Inspect the actual tool invocation and answer.
If tools appear but record access fails, check the selected organization, token expiry and content permission. If no tools appear, check the endpoint and HTTP transport. A client that requires an SSE-only endpoint needs a compatible transport rather than a different business payload.
How services get exposed
A method on a registered AI service becomes reachable through MCP when it is annotated for AI use, the same mechanism behind runtime discovery. Annotating a method makes it discoverable; leaving the annotation off keeps it out of an agent's reach.
That is the control surface — what an agent can do is decided by what is annotated, not by what the client asks for.

