We’re excited to share the beta release of our tool use API! Now developers can easily equip any LLM with hosted open-source tools for web search, web crawling, and maps data (+ more coming soon). Under the hood, it's powered by MCP — but that's just an implementation detail ↓

A small reminder from the presenter

.@AnthropicAI sent a "request for startups" in MCP at @aiDotEngineer - MCP servers in new domains (outside developer tools) - Make building servers easier - AI application security (see: tool poisoning)

(You can inspect and/or contribute to the CLI too if you are so inclined.)

The registry and CLI (>= v0.0.15) now support @code as a destination for installing MCP servers. Thanks to @ConnorPeet from team VS Code for contributing the change!

https://t.co/wkpNsK0wtH

[New post] 🔥A Deep Dive Into MCP and the Future of AI Tooling APIs were the internet’s first great unifier but AI models lack an equivalent. What are the use cases of MCPs today? Where are the challenges? Here's a technical deep dive & map of the current ecosystem 👇

(Talk slides are here.) https://t.co/DQy4TEydR1

It matters because the effort to turn existing REST APIs into more LLM-friendly MCP servers will get that much easier: - No new endpoints required (no /sse) - More options to deploy (serverless) https://t.co/zflKHgIGwG

.@AnthropicAI will support stateless remote MCP servers with “just HTTP” as transport. We’re quite excited about this! So we gave a “lightning talk” at an MCP meetup hosted by @paulgb @CloudflareDev yesterday.

This wouldn’t be possible without MCP and its fast-growing community. We hope to help deliver on its promise of augmenting LLMs with the “shortest path” to value. If this speaks to you, sign up for the waitlist and join the Discord while you wait 🛠️

The SOTA model changes every month. Model providers are incentivized to create lock-in by offering "walled garden" functionality. OpenTools aims to give you optionality in both your models and your tools. Our incentive is to help make your apps resilient to provider outages.

Before this, LLM tool use was really "structured outputs" Developers had the burden of writing integrations to external services. The OpenTools alternative: use tools from open-source MCP servers so you don’t have to reinvent the wheel. Before vs. after:

@QuickBooks More of "what's needed now" from the creator

The recording is live (announcement at 1:22:00):

> How does MCP do on the confetti benchmark? Glad you asked. https://t.co/ZaPUpR47gi

> What are the opportunities? More *high quality* MCP servers needed. Particularly for systems of record like @QuickBooks. Agents authenticating on behalf of their users still needs battle-testing: https://t.co/hNuPnTyrto https://t.co/9CWVMXYuBj

> Why not OpenAPI instead of MCP? Specs are too many tokens. GitHub’s OpenAPI is ~2M tokens, e.g. The “C” in MCP is about providing a model-friendly amount of context. Ideally, every tool description is written like a mini system prompt to clearly instruct how to use each tool.

> Why is MCP valuable? It eliminates the glue code people (or LLMs) would’ve otherwise needed to write to use external services via API. Example: If you’re building a coding agent that can file GitHub issues, you could handwrite that integration. Or, take the path of least

Here’s our “MCP FAQ” in response to some recent posts about MCP. If anything could be better explained or if you’re @dsp_, please chime in ↓ To start, > What is MCP? A balanced writeup we like by @genmon: https://t.co/EmLwov3Yt0 https://t.co/Z61uzWjLEV