feat: add inputs parameter to MCP tool invocation methods

[zinodynn]

[!IMPORTANT]

1. Make sure you have read our contribution guidelines
2. Ensure there is an associated issue and you have been assigned to it
3. Use the correct syntax to link this PR: Fixes #<issue number>.

Summary

• feat: Forward chat input variables to MCP server via _meta

  • Enables passing additional variables from chat request inputs to MCP servers, facilitating use cases like identity verification and permission control.

• feat: Support MCP protocol-reserved _meta parameter in tool invocations

  • Adds the _meta field to CallToolRequestParams to comply with the MCP specification, enabling transmission of metadata such as progressToken or identityToken.

These changes improve flexibility for MCP-level business logic, allowing for more fine-grained control and context-aware processing based on upstream chat request details.

References:

• MCP Spec - General Fields
• #28076, #28167

Screenshots

Before	After
sql execution in MCP cannot be authenticated	The functions can be expanded by customizing parameters

Checklist

• [ ] This change requires a documentation update, included: Dify Document
• [x] I understand that this PR may be closed in case there was no previous discussion or issues. (This doesn't apply to typos!)
• [x] I've added a test for each change that was introduced, and I tried as much as possible to make a single atomic change.
• [x] I've updated the documentation accordingly.
• [x] I ran dev/reformat(backend) and cd web && npx lint-staged(frontend) to appease the lint gods

[gemini-code-assist[bot]]

Summary of Changes

Hello @zinodynn, I'm Gemini Code Assist[^1]! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly enhances the functionality of Multi-Chain Protocol (MCP) tools by allowing the forwarding of extra parameters, referred to as 'inputs,' from chat requests. This new capability provides a robust mechanism for implementing advanced business logic within MCPs, such as user authentication and fine-grained access control, by leveraging contextual information from the initial chat request. The change improves the overall adaptability and power of the tool invocation system.

Highlights

• Enhanced MCP Tool Invocation: Introduced the capability to pass additional inputs parameters from chat requests directly to Multi-Chain Protocol (MCP) tool invocation methods.
• Improved Flexibility: This change enables more flexible and context-aware processing within MCPs, supporting use cases like request-side identity verification and granular permission control.
• Conditional Input Passing: The inputs are conditionally passed to MCP tools, ensuring that only relevant tools receive these additional parameters, and are merged conservatively to avoid overwriting existing tool parameters.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with :thumbsup: and :thumbsdown: on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

[^1]: Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution.

[zinodynn]

Reference my approach to pass input parameters from chat requests to MCP tools, as outlined below:

1. Create an identityToken using the app's input parameters, then save and publish it.
   

2. Modify the MCP server code, and receive the identityToken params:

export async function executeSqlToolHandler(
  { sql }: { sql: string }, 
  _extra: any // Additional parameters from chat request
) {
  const connector = ConnectorManager.getCurrentConnector();
  // Retrieve `input` params from the request via:
  // const identityToken = extra._meta?.identityToken;   <---  receive params
  try {
    // ... Use the `sql` parameter ...
  }
  // ...
}

3. Send a chat request (example curl command):

curl --location 'http://localhost:5001/v1/chat-messages' \
--header 'Authorization: Bearer app-xxxxxx' \
--header 'Content-Type: application/json' \
--data '{
    "inputs": {"identityToken":"eyJhbG...."},
    "query": "Query today'\''s sales orders?",
    "response_mode": "streaming",
    "user": "abc-123"
}'

[zinodynn]

@Nov1c444, to prevent every MCP from receiving the parameters, would it be a good idea to distinguish parameters in inputs using dot-separated notation (e.g., .)? For example, create a parameter like mysql-mcp-server.identityToken in inputs to pass it specifically to the mysql-mcp-server MCP. This way, other MCPs won’t receive the identityToken parameter. If this makes sense, I’ll proceed to work on it.

[asukaminato0721]

/gemini review

[Nov1c444]

Hey there! 👋 I understand this PR aims to solve a core problem: allowing MCP tools to receive request-level dynamic context information, including:

• User identity information (roles, IDs, etc.) for business logic differentiation (ref #28167)
• Authentication credentials (access tokens, etc.) for MCP server authentication (ref #28076) The current implementation passes this data to MCP servers as _meta via the inputs parameter, which indeed solves the problem in Agent Apps 👍 However, there are still a few issues we need to optimize:

1. Workflow/Chatflow cannot use the same mechanism The current solution only works in Agent Apps. If users want to invoke MCP tools based on user roles in Workflows, or pass dynamic tokens, there's no corresponding solution. The same MCP tool has inconsistent capabilities across different application types, which creates a fragmented user experience.
2. Using reflection mechanism lacks clarity The base class uses inspect.signature() to detect if subclasses accept the inputs parameter:

 sig = inspect.signature(self._invoke)   
 if "inputs" in sig.parameters:       
    invoke_kwargs["inputs"] = inputs

While this works, the interface contract is unclear, IDEs can't provide type hints, and it's not very developer-friendly for tool authors. 3. Cannot directly pass to HTTP Headers (for #28076) Currently _meta is passed in the request body. If users want to put tokens in HTTP request headers (like IDToken: xxx), they still need to manually handle it on the MCP server side, which isn't straightforward. I need to think through the specific implementation approach and will update later. Welcome everyone to discuss.

[Nov1c444]

I have some thoughts about this requirement. Add a meta configuration to the MCP Server, like this. Does this design solve the issue? If so, we can discuss more details.:

1. When clicking the MCP tool settings, you could see two tabs: Parameters and Meta
2. Parameters are LLM-generated, while Meta is user-defined
3. You could add some meta key-value pairs in the Meta tab. The meta values could be variables that you define yourself, and they will be passed into the MCP meta field.
4. Agent apps, workflows, and chatflows would work similarly.

Before	After

[zinodynn]

• When clicking the MCP tool settings, you could see two tabs: Parameters and Meta
• Parameters are LLM-generated, while Meta is user-defined
• You could add some meta key-value pairs in the Meta tab. The meta values could be variables that you define yourself, and they will be passed into the MCP meta field.
• Agent apps, workflows, and chatflows would work similarly.

Totally agree with the new changes you proposed! Honestly, I didn’t realize the specificity of the inputs parameter in chat requests before. I kind of had a gut feeling that MCP-related configurations should be set in the MCP tool’s own properties, but I still went with inputs at the time—since I thought this parameter persists throughout the entire chat flow, making it easier to fetch values. If we go with the Meta tab design you mentioned, I’m curious how we’d pass these values to the actual HTTP requests? From what I know, request parameters often get filtered. For example, would something like IDkey still need to be included in inputs to be sent in the request?

[Nov1c444]

I mean, you could pass the IDKey variable through the inputs, like this:

[zinodynn]

So does that mean we’ll still pass parameters via inputs, but the configuration and metadata will be handled differently? @Nov1c444

[Nov1c444]

So does that mean we’ll still pass parameters via inputs, but the configuration and metadata will be handled differently? @Nov1c444

Yeah, it's the best way I could think of to pass the variable in the Agent App's MCP Tool. Do you have any better ideas?

[zinodynn]

@Nov1c444 Hi , Here are my understandings and follow-up questions:

1. If we stick with passing parameters via inputs, after configuring the key in MCP Meta, will the Workflow fetch the specific value from inputs using that key? Due to Dify's parameter filtering mechanism, it seems like this would require two configurations: first, adding the corresponding field in inputs, and second, creating a field in MCP Meta where the key matches the one in inputs (the value can either be passed from inputs or set as a default). Is that correct?

2. For passing parameters like IDKey via HTTP headers—what would the workflow look like? Do we just need to create a field in MCP Meta? Will Dify filter out these header parameters, or will they be forwarded directly to the MCP server?

[zinodynn]

In this example, the Field value should be inputs.IDkey. right? id_key hasn't been defined.

[zinodynn]

Headers and _meta are two separate channels: headers work at the HTTP layer, while _meta operates at the MCP protocol layer. I’m wondering if defining something like IDkey in headers is appropriate—after all, in the current code, headers are mainly used for OAuth authentication. I think it’s better to keep it within the MCP protocol layer, i.e., defining it in inputs. Looking forward to everyone's thoughts/discussions!

[cion-J]

If I make a request through the API, such as /workflows/run, and customize the header in the request, can it be attached to the MCP request? inputs will disrupt the parameter model of the mcp service.

[YQP-2511]

Is it the difference between active reception and passive reception parameters of the MCP service? In Field VALUE, we can use / to select parameters defined in AGENT/CHATFLOW/WORKFLOW just like in INSTRUCTIONS. Are these applications required to use that MCP in order to select through / in the _meta of that MCP?

[cion-J]

The situation I encountered is that there is no need to configure or adjust each MCP individually. If this global _meta or variable has been determined at the beginning, it can be automatically passed down to each MCP, just like tracking. This will reduce the workload when configuring multiple MCPs.

[tvp100]

I need this feature to set meta data, so I can get it from McpMeta. this is my java code, use spring ai mcp

    @McpTool(name = "getAreaTree", description = "获取区域相关的信息, 能管的设备分4个, 电、水、气、秤, 返回值的areaIdChain表示从顶级区域开始到当前区域的ID链, 用'/'分割")
    public List<AreaTreeNodeDto> getAreaTree(McpSyncRequestContext context, AiListAreaTreeParam param, McpMeta meta) {
        return remoteSystemOpenService.getAreaTree(param).getData();
    }

and this is another tool which has the feature I need.

is there anyway I cloud pass it through Agent in dify? or will this feature be merged?

In this example, the `Field value` should be `inputs.IDkey`. right? `id_key` hasn't been defined.