[Bug]: Composio MCP tool-call input sometimes emitted as string, breaks AI SDK tool_use validation

[kdawgwilk]

SDK Language

TypeScript / Node.js SDK (@composio/core)

SDK Version

@composio/core@0.5.1

Runtime Environment

Node.js 24 on macOS 15.7.2

Environment

Production Deployment

Describe the Bug

Summary
We’re seeing Composio MCP tool calls occasionally emit input as a string instead of an object. When using Vercel AI SDK v5 streaming, this causes a runtime validation error (tool_use.input: Input should be a valid dictionary) and breaks chat streaming.

Environment

• AI SDK: ai@5.x
• Composio: @composio/core + @composio/vercel (current versions at time of issue)
• Integration: AI SDK streamText() + toUIMessageStream()
• Tool: COMPOSIO_MULTI_EXECUTE_TOOL

Actual behavior
AI SDK receives a tool-call with input as a string (often JSON-looking but not parseable), which fails its schema validation and terminates streaming.

Expected behavior
Tool-call input should always be a JSON object (or valid structured data) consistent with AI SDK’s tool call schema.

Example (sanitized)
toolName: COMPOSIO_MULTI_EXECUTE_TOOL
input type: string (length ~400)
Parsing fails or produces a non-object value.

Impact

• AI SDK stream fails with validation error
• UI shows generic chat error
• Requires manual normalization to avoid crashes

Notes
We implemented a temporary server-side normalization (coerce string → {} or parsed object) to prevent crashes. Ideally Composio should emit object-shaped tool input compatible with AI SDK expectations.

Steps to Reproduce

1. Configure Composio MCP tools using @composio/core + @composio/vercel and create an AI SDK streamText() call with those tools enabled.
2. Trigger a request that uses COMPOSIO_MULTI_EXECUTE_TOOL (e.g., ask to send a Slack DM).
3. Observe the streaming response fail with tool input validation.

Minimal Reproducible Example

ts
import { streamText } from 'ai';
import { createComposioMcpTools } from './composio-mcp';
const { tools } = await createComposioMcpTools('user-123');
const result = streamText({
model: /* any AI SDK model /,
messages: [{ role: 'user', content: 'Send a Slack message to Someone' }],
tools,
});
for await (const chunk of result.toUIMessageStream()) {
// Stream to client
}

Error Output / Stack Trace

messages.3.content.1.tool_use.input: Input should be a valid dictionary

Reproducibility

• Always reproducible
• Intermittent / Sometimes
• Happened once, can’t reproduce

Additional Context or Screenshots

No response

[Uday-sidagana]

Hey @kdawgwilk, could you share more details about how you are handling tools( createComposioMcpTools )? Also, please share the model you are using.

[kdawgwilk]

Yes using the createComposioMcpTools and I was seeing this with claude-haiku-4-5