Summary

When using MCP tools with strict schemas (e.g. { "text": "string" }), LibreChat forwards malformed tool arguments produced by the LLM without validation or normalization.

This leads to downstream failures such as:

Error: Received tool input did not match expected schema

Environment

• LibreChat (latest)
• LangChain / LangGraph orchestration
• MCP servers using Snowflake CORTEX_AGENT_RUN
• Model: Claude Sonnet 4.5 (Anthropic)

Reproduction

User prompt (multi-intent)

use grdt-tools to answer: gill mortality for last year
use bi-tools to answer: how can I get access to snowflake?

Observed LLM tool call output

{
  "tool_calls": [
    {
      "function": {
        "arguments": "{\"text\": \"What was the gill mortality for last year?\"}{\"text\": \"How can I get access to Snowflake?\"}"
      }
    }
  ]
}

Problem

The arguments field contains:

{...}{...}

This is:

• not valid JSON
• not an array
• not parseable

LibreChat forwards this as-is to the MCP server, which rejects it due to schema mismatch.

Expected Behavior

LibreChat should not blindly forward malformed tool arguments.

Instead, it should:

Option A (preferred)

Detect concatenated JSON objects and split:

{...}{...} → [{...}, {...}]

Then:

• execute sequentially OR
• pass individually to tools

Option B

Attempt safe normalization:

{...}{...} → { "text": "A\nB" }

Option C

Fail early with a clear validation error, e.g.:

Invalid tool arguments: expected valid JSON object, got concatenated JSON

Why this matters

• This is a common failure mode with multi-intent prompts and modern LLMs
• Particularly frequent with Anthropic models
• Breaks MCP integrations with strict schemas
• Currently requires external proxies or workarounds

LibreChat is the ideal layer to handle this safely.

Suggested Implementation

At tool execution boundary:

1. Validate arguments as JSON

2. If invalid:

   • detect {...}{...} pattern (regex)
   • split into multiple JSON objects

3. Either:

   • execute sequentially
   • or normalize into a single valid payload

Proposal

Add a tool argument validation + normalization layer before MCP execution.

This would:

• improve robustness across all LLM providers
• prevent downstream schema errors
• reduce need for custom middleware

Thanks!