No matter how fast we shipped MCP tools, it wasn't fast enough. So we let agents build them. Meet Gauntlet, the loop that produced 10 new toolkits in a weekend, work that used to take two months.
Sending a file through an AI agent shouldn't cost a million tokens or a file picker. Here's how Arcade.dev keeps file bytes out of the model's context, and why MCP's file story needs a client-resolution mode before autonomous agents hit a wall.
TL;DR: Anthropic's new Tool Search is a step in the right direction-but if you're running 4,000+ tools across multiple services, it might not be ready for prime time. The promise Anthropic's Tool Search promises to let Claude "access thousands of tools without consuming its context window." Music to our ears. At Arcade, we maintain thousands of agent-optimized tools across Gmail, Slack, GitHub, HubSpot, Salesforce, and dozens more platforms. If anyone was going to stress-test this feature, it
Your MCP server works perfectly on localhost. Five green checkmarks. Clean logs. You're a genius! Then you try to deploy it. OAuth tokens leak into logs. Secrets are hardcoded because "it's just a prototype." The whole thing crashes under two concurrent users. Welcome to production. Every developer hits this wall. You can build an MCP server in an afternoon, but making it production-ready usually means rewriting half of it. New auth flows. Proper secret management. Multi-user context handlin
