Invalid schema for function hubspot-search-objects

[clemenspeters]

When using MCPServerStdio with @openai/agents I get

error: {
message: "Invalid schema for function 'hubspot_search_objects': In context=('properties', 'filterGroups', 'items', 'properties', 'filters', 'items', 'properties', 'values'), array schema missing items.",
type: 'invalid_request_error',
param: 'tools[2].parameters',
code: 'invalid_function_parameters'
}

Current workaround:

    toolFilter: createMCPToolStaticFilter({
        blocked: ['hubspot-search-objects'],
    }),

fixes the issue by ignoring that function

[Paul-EA-3]

I ran into this exact same schema validation error when using the HubSpot MCP server with Gemini.

The root cause of the missing field error is in src/tools/objects/searchObjectsTool.ts. The
schema uses z.any() for several fields (like values, value, and highValue), which causes Zod to
generate a JSON schema without a type definition for the array items. Strict AI agent frameworks
(like Gemini's GenerateContentRequest and OpenAI) reject arrays that don't explicitly define their
item types.

The Fix:
In searchObjectsTool.ts (around line 24-35 in the compiled output), replacing the instances of z.any() with z.string() completely resolves the issue and generates a valid JSON schema.

    1 > const FilterSchema = z.object({
    2 >     propertyName: z.string().describe('The name of the property to filter by'),
    3 >     operator: OperatorEnum.describe('The operator to use for comparison'),
    4 >     // CHANGE: value: z.any() -> value: z.string()
    5 >     value: z.string().describe('The value to compare against. Must be a string'),
    6 >     values: z
    7 >         // CHANGE: .array(z.any()) -> .array(z.string())
    8 >         .array(z.string())
    9 >         .optional()
   10 >         .describe('Set of string values for multi-value operators like IN and NOT_IN.'),
   11 >     highValue: z
   12 >         // CHANGE: .any() -> .string()
   13 >         .string()
   14 >         .optional()
   15 >         .describe('The upper bound value for range operators like BETWEEN. The lower bound is
      specified by the value attribute'),
   16 > });
   17 > 

I patched this locally in my node_modules and it immediately fixed the HTTP 400 validation errors, allowing the search-objects tool to work perfectly!

[ross-makeacompany]

Confirmed reproducible on @hubspot/mcp-server@0.4.0 against any strict-schema MCP client. The schema returned by tools/list for hubspot-search-objects includes:

"filters": {
  "type": "array",
  "items": {
    "type": "object",
    "properties": {
      "values": { "type": "array" }   // <-- missing "items"
    }
  }
}

OpenAI's function-calling validator (and any other strict JSON-Schema validator) rejects this — arrays must declare items. Claude's clients happen to be permissive about it so the issue is invisible there, which is probably why it shipped.

For folks hitting this and unable to use toolFilter: in the meantime we worked around it by reimplementing only the tools we actually needed against HubSpot's REST API directly, behind our own MCP server. Schema gets emitted from our handlers, so we know it's valid. Painful, but unblocked us.

Real fix is one-line on your side — add "items": { "type": "string" } (or whatever HubSpot's filter values actually accepts) to the values array schema. Would happily PR if #1 ever lands.