AnQiCMS v3.6.5 Added MCP intent directory and round-by-round batch approval, how to connect external AI client

What new features does the AnQiCMS v3.6.5 version MCP Server support, and how to integrate external AI tools like Claude, Cursor?v3.6.5 released on October 7, 2026, this version rebuilds the MCP tool layer based on the intent directory and adds round-level batch approval: multiple write operations in the same round of conversation will be merged into one confirmation, instead of popping up interruptively one by one.The access method continues to use Bearer Token authentication. After configuring the token, external AI clients such as Claude, Cursor, etc. that support MCP can call the site capabilities.

The Intent Directory aggregates the tool interface into a set of intents

The early MCP access method was close to mapping each backend interface to a tool individually, and as the number of tools expanded, the probability of selecting the wrong tool and making repeated calls also increased. The approach of v3.6.5 is to converge on intent: the tool interface is uniformly organized based on the intent catalog, with each intent corresponding to a business matter such as creating articles, checking categories, and uploading images, rather than the original shape of the interface. For AI clients, there are fewer options and clearer semantics, resulting in a naturally higher success rate in composition; for site administrators, permissions and audits can also be viewed by intent domain rather than scattered across dozens of interfaces.

The round-level batch approval solves the problem of too many interruptions.

Batch approval is another highlight of this version. AI often performs multiple write operations in a single task, and confirming each one can chop up the interaction. The round-level approval approach is to merge and confirm write operations within the same round: the client initiates a batch of actions, which are summarized into a single check on the approval interface, and after seeing the scope of impact, approve or reject once. Read operations do not occupy the approval quota, and write operations must be confirmed, with the rhythm of both sides returning to their respective places. The accompanying email subscription, message list, attachment base64 upload, and other capabilities have also entered the tool layer in the same version.

Pre-access authorization and exposure setting

MCP Server uses Bearer Token authentication, the token is generated in the background, and it is recommended to hold it separately by environment and rotate it regularly; when upgrading to a new version, it is a common practice to rotate the site signature key and change the administrator password, and similar suggestions are also given in the v3.6.6 fix notes. In terms of exposure, the interface is set to a visible range according to the intent domain, and high-risk domains such as backup, upgrade, and multi-site are closed by default (DefaultOff), and need to be explicitly enabled - if only external clients are allowed to maintain content, the high-risk domains can remain closed.

Steps for integrating Claude, Cursor

Step 1, confirm that the site has been upgraded to version 3.6.5 or above and the MCP Server has been enabled in the background; Step 2, generate the Bearer Token, only hand it over to the environments that need to use it; Step 3, fill in the service address and token in the MCP configuration of Claude or Cursor; Step 4, run a read-only intent (such as checking the document list, checking categories) to verify connectivity first, and then open the content intent as needed. The entire process does not require installing any additional plugins on the site, and the client-side configuration is complete.

Typical usage and boundaries

Common usage after connection includes three types: on the content side, let AI batch draft, add images, and classify according to选题; on the operation side, use natural language to check messages and organize tags; on the development side, entrust the interface debugging and template checking to the client agent. It is also necessary to clarify the boundaries: high-risk areas should not be enabled, so AI cannot touch backup and upgrade; write operations that have not passed round confirmation will not be stored - this is a design guardrail, not a defect.

Common questions

Has there been a change in the connection method after v3.6.5?The change is in the organization method at the tool layer, authentication is still Bearer Token, and the client configuration items remain unchanged.

Will batch approval amplify accidental operations?The merged is an affirmation action, not a validation logic; each write operation is still visible item by item in the summary list.

Can it handle multiple clients at the same time?Yes, tokens can be issued multiple times for different purposes, making it convenient for separate auditing and revocation.

What should be checked first if Claude Desktop cannot connect?First, check the external access address and certificate configuration; then, test connectivity with the read intent; finally, check the token permission domain.