Hi folk,
I have a local mcp server that interacts with a database. The mcp has several tool functions, some of them are in read others in write.
I want to be able to:
- or approval for all mcp tools
- or approval only for write operations (or more in general way: per tool function approval).
What is working
mcpServer := agents.NewMCPServerStdio(agents.MCPServerStdioParams{
Name: "mcp-mysql",
Command: cmd,
CacheToolsList: true,
})
agent.
AddMCPServer(mcpServer)
With this code the agent is able to use my mcp server without any issue but it executes the tool functions without asking for permission.
Looking at the code the only place I found something similar is hosted_mcp: https://github.com/nlpodyssey/openai-agents-go/blob/main/examples/hosted_mcp/approvals/approvals.go
I tired to use it in combo with mcp server but it seems that the hostedMcp methos is designed to work only with remote mcp servers
mcpServer := agents.NewMCPServerStdio(agents.MCPServerStdioParams{
Name: "mcp-mysql",
Command: cmd,
CacheToolsList: true,
})
agent.
AddMCPServer(mcpServer).
WithTools(
agents.HostedMCPTool{
ToolConfig: responses.ToolMcpParam{
ServerLabel: "mcp-mysql", <--- same name as the mcpServer instance
ConnectorID: "asd", <--- tried different values or ServerURL, without any luck
RequireApproval: responses.ToolMcpRequireApprovalUnionParam{
OfMcpToolApprovalSetting: param.NewOpt("always"),
},
Type: constant.ValueOf[constant.Mcp](),
},
OnApprovalRequest: approvalCallback,
},
).
I looked around and found also: hooks, however they seem to be designed as "notification" system of events and not as control mechanism.
Then I started to dig into the code and debugged the flow:
if my understanding is correct to cover my use-case something is missing in the mcp_util.go file method InvokeMCPTool. Or maybe better extend the FunctionTool struct to support a tool approval callback.
Is there an already supported way to implement my use-case?
Thank you for your support
Hi folk,
I have a local mcp server that interacts with a database. The mcp has several tool functions, some of them are in read others in write.
I want to be able to:
What is working
With this code the agent is able to use my mcp server without any issue but it executes the tool functions without asking for permission.
Looking at the code the only place I found something similar is hosted_mcp: https://github.com/nlpodyssey/openai-agents-go/blob/main/examples/hosted_mcp/approvals/approvals.go
I tired to use it in combo with mcp server but it seems that the hostedMcp methos is designed to work only with remote mcp servers
I looked around and found also: hooks, however they seem to be designed as "notification" system of events and not as control mechanism.
Then I started to dig into the code and debugged the flow:
if my understanding is correct to cover my use-case something is missing in the mcp_util.go file method InvokeMCPTool. Or maybe better extend the FunctionTool struct to support a tool approval callback.
Is there an already supported way to implement my use-case?
Thank you for your support