MCP Settings
The MCP settings control the MCP server that QIE hosts on its own port. They are stored once for the whole installation, so every node in a high availability cluster uses the same values. Administrators change them on the Model Context Protocol (MCP) page in System Administration. An AI client connected through the MCP can read them but never change them. A Remote Management Hub has MCP settings of its own, on its own page, without Return message content and Require client confirmation (see The MCP Server on a Hub).
Every change is recorded in the revision history, like any other configuration change. The settings travel with System Configuration in a configuration export and import, so restoring an environment from an export restores its MCP settings too, Enable the MCP server included. When the restored settings turn the server on, the listener starts if its port and certificate can be used on that installation, and is tried again every minute if not (see MCP Listener). A Hub exports and restores its own MCP settings the same way. Registered clients and sign-ins never travel: after a restore to a new installation, each person approves their AI client again. The server certificate and trusted client certificates travel as references to Global Settings certificates, and one the import does not find among the Global Settings certificates is left empty, so export the Global Settings certificates with them.
MCP Server
| Setting |
Default |
Effect |
| Enable the MCP server |
Off |
Turns the MCP server on. While it is off, the listener does not run and every MCP sign-in address on the console answers as if it did not exist. |
| MCP Port |
None |
The port the listener uses. Required before the server can be enabled. It must differ from the console port, from the Management API port while the Management API is on, and from every port a channel listens on. On a Hub it must also differ from the tunnel listener's Port in Hub Configuration. |
| Server Certificate |
None |
The certificate, with its private key, that the listener presents. Choose one from Global Settings. Required before the server can be enabled, because the listener accepts HTTPS only. |
| Issuer URL |
The console address followed by /qie |
The address clients use to reach the console sign-in. It must be an HTTPS address. |
| Resource URL |
https://<server name>:<MCP port>/mcp |
The address clients use to reach the listener. Every access token is issued for this address only, so changing it signs every client out. When the server's name is not a valid host name, such as one with an underscore, enter the address here: the MCP server cannot be turned on without one. |
| Environment Label |
Empty |
A short name such as PROD or DEV that clients show beside the server name, so a user can tell two engines apart. Letters, digits, spaces, hyphens and underscores, up to 20 characters. |
Client Certificates
| Setting |
Default |
Effect |
| Client Auth |
Off |
Require a certificate from every client asks every client for a certificate and accepts only certificates issued by one of the trusted authorities. |
| Trust Policy |
Trust Selected Certificates |
The certificates ticked in the list below it are the trusted authorities for client certificates, used while Client Auth is on. None is ticked by default. |
Content and Confirmation
| Setting |
Default |
Effect |
| Return message content |
Off |
Allows the MCP to return message bodies, log text, cache values and sample messages. While it is off, those values are left out of every result. |
| Require client confirmation |
On |
Refuses the irreversible tools, such as deletes, purges and imports, to a client that cannot ask its user to confirm, and refuses a report to Qvera from such a client. The user can still make those changes in the console. |
| Allow reports to Qvera |
On |
Lists the send_report tool, through which a person sends a report about an MCP session to Qvera support: the model's account, the person's comment and the session's audit entries with message content left out. The person approves each report in a form before it is sent. While Require client confirmation is off, a client that cannot show a form passes on the person's answer instead. See Sending a report to Qvera. |
Tokens
| Setting |
Default |
Effect |
| Access Token Lifetime (seconds) |
3600 |
How long an access token lasts, from 300 to 86400 seconds. |
| Issue refresh tokens |
On |
Lets a client renew its access token without the user signing in again. Signing out of the console does not end a client's tokens, and a refresh token skips two-factor authentication for its lifetime. |
| Local User Refresh Lifetime (days) |
30 |
The longest a client stays signed in for a user whose password QIE checks, from 1 to 365 days. |
| LDAP and OIDC User Refresh Lifetime (days) |
1 |
The longest a client stays signed in for a user checked by LDAP or an OIDC provider, from 1 to 365 days. The default is shorter because QIE cannot check the directory again when a client renews its token. |
| Metadata Document Hosts |
Empty |
Hosts allowed to publish client metadata documents, one per line, such as claude.ai for Claude Code or chatgpt.com for Codex. A host name on its own allows port 443 only; add :port for another port. While the list is empty, QIE fetches no document and clients register directly instead (see MCP Client Sign-In). |
Limits on Each Node
Each limit applies on each node. A cluster behind a load balancer therefore allows the setting multiplied by the number of nodes. For calls through the Hub, an engine applies Requests per Token per Minute, Concurrent Requests per Token and Message Searches per Token per Minute to each person the Hub vouches for, and refuses a call over Concurrent Requests per Token at once (see MCP Calls Through the Hub).
| Setting |
Default |
Effect |
| Requests per Token per Minute |
120 |
Calls one token may make each minute. |
| Concurrent Requests per Token |
8 |
Tool calls and resource reads one sign-in may have running at once. Another one waits up to 2 seconds for a place before it is refused. |
| Message Searches per Token per Minute |
10 |
Message searches one token may start each minute. |
| Unauthenticated Requests per Address per Minute |
60 |
Requests without a token that one address may make to the listener each minute. |
| Registrations per Address per Hour |
10 |
Client registrations one address may make each hour. |
| Token Requests per Address per Minute |
30 |
Token, revocation and authorization requests one address may make each minute, each kind counted on its own. |
| Max Page Size |
100 |
The most items any list returns at once. |
| Worker Threads |
16 |
Threads the listener uses to receive requests. Tool calls run on a separate pool the MCP library manages, so this does not cap how many run at once; Concurrent Requests per Token limits that for each sign-in. |
| Request Timeout (seconds) |
60 |
How long the MCP waits for a person to answer a confirmation, for an import before it returns a status to poll, and for the published-mapping usage check. Other calls are not cut off. |
| Message Search Timeout (seconds) |
30 |
How long the MCP waits for a message search. The database may keep working on the search after the MCP stops waiting. |
| Read Cap (bytes) |
40000 |
The largest result one call returns. A longer result is cut off with a marker, and the client can ask for the rest by range. |
| Package Ceiling (bytes) |
10485760 |
The largest configuration export the MCP returns and the largest import it accepts. |
Audit Log
| Setting |
Default |
Effect |
| Audit Retention (days) |
365 |
How long the MCP audit log keeps its entries. 0 keeps them for ever; any other value is at least 30 days. On an engine registered with a Hub, it is also how long the engine keeps the statements the Hub sends. |