MCP - Handling Complex Schemas + System Prompt

[lnvestor]

Claude/Anthropic Model Unable to Handle Complex JSON Schema with anyOf/oneOf

Description

I've set up LibreChat with a custom MCP server containing 20 tools with complex schemas. While the tools are correctly displayed and accessible, I'm encountering issues when Claude (Anthropic's model) attempts to create connections using schemas that contain anyOf/oneOf conditions.https://github.com/danny-avila/LibreChat/issues/new/choose

Environment

• LibreChat (Docker deployment)
• Custom MCP server
• Anthropic's latest model with Tools support

Configuration

docker-compose.override.yml

services:
  api:
    environment:
      - MONGO_URI=${MONGO_URI}
    volumes:
      - ./librechat.yaml:/app/librechat.yaml
      - /Users/drisstalainte/Documents/Cline/MCP:/app/mcp
  mongodb:
    image: arm64v8/mongo:latest
    command: ""
    entrypoint: ""

MCP server yaml config

myMCP:
  type: stdio
  command: "node"
  args:
    - "/app/mcp/MyMCP-server/build/index.js"
  env:
    NODE_ENV: production
    MyMCP_BEARER_TOKEN: "f596**********9dcf7f44c0"

Error

When attempting to create a connection, the following error occurs:

2025-01-24 00:05:26 LibreChat | 2025-01-23 23:05:26 error: create_connection MCP server tool call failed MCP error -32603: MCP error -32603: MCP error -32602: Missing required parameters: type, name

Investigation

The issue appears to be related to how complex JSON schemas with anyOf/oneOf conditions are handled. Looking at Cline's implementation (specifically this code), they handle this by:

1. Using custom prompts
2. Using JSON.stringify to send the input schema

Potential Solution

Implementing similar handling of complex schemas as Cline does might resolve this issue.

Additional Context

• Full schema implementation is provided in the comment below (due to length)
• Screenshot of tools display

---

Would you like me to post the schema as a comment to maintain better readability of the main issue?

[lnvestor]

export const httpConnectionSchema = { type: 'object', properties: { type: { type: 'string', const: 'http', description: 'Connection type', }, name: { type: 'string', description: 'Connection name', }, offline: { type: 'boolean', description: 'Whether the connection is offline', }, sandbox: { type: 'boolean', description: 'Whether to use sandbox environment', }, http: { type: 'object', properties: { formType: { type: 'string', enum: ['http', 'graph_ql'], description: 'Form type for HTTP connections', }, mediaType: { type: 'string', enum: ['json'], description: 'Media type', }, baseURI: { type: 'string', description: 'Base URI for the endpoint', }, unencrypted: { type: 'object', properties: { field: { type: 'string', description: 'Unencrypted field value', }, }, }, encrypted: { type: 'string', description: 'Encrypted data', }, auth: { type: 'object', properties: { type: { type: 'string', enum: ['basic', 'cookie', 'digest', 'token'], description: 'Authentication type', }, basic: { type: 'object', properties: { username: { type: 'string', description: 'Basic/Digest auth username', }, password: { type: 'string', description: 'Basic/Digest auth password', }, }, required: ['username', 'password'], }, cookie: { type: 'object', properties: { uri: { type: 'string', description: 'Cookie auth URI', }, method: { type: 'string', description: 'HTTP method for cookie auth', enum: ['GET', 'POST'], }, }, required: ['uri', 'method'], }, token: { type: 'object', properties: { token: { type: 'string', description: 'Authentication token', }, location: { type: 'string', enum: ['body', 'header', 'url'], description: 'Token location (body, header, or url)', }, headerName: { type: 'string', description: 'Header name when using header location', }, scheme: { type: 'string', enum: ['Bearer', 'custom', 'mac', 'None', ' '], description: 'Authentication scheme (Bearer, custom, mac, None, or space for no scheme)', }, paramName: { type: 'string', description: 'Parameter name for token', }, }, required: ['token', 'location', 'headerName', 'scheme', 'paramName'], allOf: [ { if: { properties: { location: { const: 'header' } } }, then: { properties: { scheme: { enum: ['Bearer', 'custom', 'mac', 'None'], }, }, } }, { if: { properties: { location: { const: 'body' } } }, then: { properties: { scheme: { const: ' ', }, }, } } ], }, }, required: ['type'], allOf: [ { if: { properties: { type: { const: 'basic' } } }, then: { required: ['basic'] } }, { if: { properties: { type: { const: 'cookie' } } }, then: { required: ['cookie'] } }, { if: { properties: { type: { const: 'digest' } } }, then: { required: ['basic'] } }, { if: { properties: { type: { const: 'token' } } }, then: { required: ['token'] } } ], }, ping: { type: 'object', properties: { relativeURI: { type: 'string', description: 'Relative URI for ping endpoint', }, method: { type: 'string', description: 'HTTP method for ping', }, }, }, }, required: ['formType', 'mediaType', 'baseURI', 'auth'], }, microServices: { type: 'object', properties: { disableNetSuiteWebServices: { type: 'boolean', default: false, }, disableRdbms: { type: 'boolean', default: false, }, disableDataWarehouse: { type: 'boolean', default: false, }, }, required: ['disableNetSuiteWebServices', 'disableRdbms', 'disableDataWarehouse'], }, queues: { type: 'array', items: { type: 'object', properties: { name: { type: 'string', description: 'Queue name', }, size: { type: 'number', description: 'Queue size', }, }, required: ['name', 'size'], }, }, }, required: ['type', 'name', 'http', 'microServices'], };

[dvejsada]

Is it possible this may be caused by using "type" as property name, when "type" is a reserved keyword in JSON Schema used to define the data type of a property?

[lnvestor]

@dvejsada i have no idea i mean in cline it works fine but running same prompt in librechat it throwing this error

[danny-avila]

Not experiencing this

"parameters": {
  "anyOf": [
    {
      "type": "object",
      "properties": {
      },
      "additionalProperties": false
    },
    {
      "type": "null"
    }
  ],
  "description": "Optional parameters required for connection. Not needed if using a default connector. Examples: {'api_key': '1234567890'}"
}

also firecrawl_crawl has this kind of property on the webhook field

"webhook": {
  "oneOf": [
    {
      "type": "string",
      "description": "Webhook URL to notify when crawl is complete"
    },
    {
      "type": "object",
      "properties": {
        "url": {
          "type": "string",
          "description": "Webhook URL"
        },
        "headers": {
          "type": "object",
          "description": "Custom headers for webhook requests"
        }
      },
      "required": [
        "url"
      ]
    }
  ]
}

[Daniel-Uzcategui]

having the same problem GenerateContentRequest.tools[0].function_declarations[8].parameters.properties[newValue].type: must be specified when not using one_of, any_of or all_of fields on my end seems to be ignoring the declaration of anyOf,oneOf, etc. Works perflectly on cline

[danny-avila]

anyOf,oneOf are not supported by Gemini API: https://discuss.ai.google.dev/t/oneof-in-response-schema/55926/46

Since you mentioned GenerateContentRequest:
https://cloud.google.com/vertex-ai/generative-ai/docs/reference/java/latest/com.google.cloud.vertexai.api.GenerateContentRequest

No issues with Anthropic. Cline likely discards those properties, may have to do that, but would prefer if the API were not strict about it or supported it.

[Daniel-Uzcategui]

anyOf,oneOf are not supported by Gemini API: https://discuss.ai.google.dev/t/oneof-in-response-schema/55926/46

Since you mentioned GenerateContentRequest: https://cloud.google.com/vertex-ai/generative-ai/docs/reference/java/latest/com.google.cloud.vertexai.api.GenerateContentRequest

No issues with Anthropic. Cline likely discards those properties, may have to do that, but would prefer if the API were not strict about it or supported it.

Oh sorry about that, my mistake

[ezyang]

It seems better to then just drop the schema info entirely. If the model passes in something wrong (which it won't, because the tool description will steer it the right way) you can just give an error saying what failed and the model will try again.

[danny-avila]

Handling here: #7040