[Audit] MEDIUM: No timeout on MCP tool calls

[NickCalabs]

Severity: MEDIUM

Description

The callTool function in src/tools/registry.ts:115-128 calls MCP tools without any timeout, allowing hung or slow tools to block agent execution indefinitely.

Location

File: src/tools/registry.ts
Line: 127

Problematic Code

return server.client!.callTool(originalName, args);

No timeout is set on the MCP client tool call operation.

Issues

1. Infinite Hangs: A buggy MCP server can hang forever
2. Resource Exhaustion: Agent threads blocked waiting for tools
3. Cascading Failures: One slow tool blocks entire agent run
4. Poor UX: Users see agents "running" but they're just waiting

Impact

• Agent runs never complete
• Database grows with stuck "running" entries
• Server becomes unresponsive over time
• No way to recover without restart

Recommendation

Add timeout to tool calls:

export async function callTool(
  toolName: string,
  args?: Record<string, unknown>,
  timeoutMs: number = 60_000, // 60 second default
): Promise<unknown> {
  const { server, originalName } = getServerForTool(toolName);

  // Local handler (no MCP subprocess)
  if (server.localHandlers) {
    const handler = server.localHandlers.get(originalName);
    if (handler) {
      // Wrap in timeout
      return Promise.race([
        handler(args ?? {}),
        new Promise((_, reject) => 
          setTimeout(() => reject(new Error(`Tool call timed out after ${timeoutMs}ms`)), timeoutMs)
        ),
      ]);
    }
  }

  // MCP client call with timeout
  return Promise.race([
    server.client!.callTool(originalName, args),
    new Promise((_, reject) => 
      setTimeout(() => reject(new Error(`MCP tool call timed out after ${timeoutMs}ms`)), timeoutMs)
    ),
  ]);
}

Also consider:

1. Making timeout configurable per-tool or per-agent
2. Logging slow tool calls (e.g., >5s) as warnings
3. Adding metrics for tool call durations

Related Code

• runner.ts doesn't have explicit timeout handling for tool calls either
• Only ollama.ts:7 sets OLLAMA_TIMEOUT_MS = 10 * 60 * 1000 for LLM calls

---

Created by security audit