Transcription
We've heard a lot from Anthropic recently, specifically about MCP tools. And with the launch of Opus 4.5, they also came at us with a really interesting solution to the MCP problem, which will go over some actual examples on how to use this new tool.
Now, if you don't know what MCPS are, they're basically in a nutshell APIs that the claw code and all these AIs can use to expand on their functionality. MCPS can access linear for your project management. It can update Jira tickets. It can search documentation all with natural language. Really powerful.
But the downside is it consumes so much context. So if I just spin up claude here and hit context, I've got linear server running which is gosh don't know how many ref x ref and exa here which are document and internet searching things. Basically my MTP tools are taking up 20,000 tokens which is 10% of my overall context. It's ridiculous.
So with the launch of Opus 4.5, they also told us about a new tool that they've released. This is kind of aimed at enterprise people building tools and agents for their system because it's available as a beta with the anthropic API. It's not yet available directly inside of clawed code. It's not something we can really use as end users. No doubt they'll refine this as they start to learn how people use it and it will eventually end it end up inside of clawed code.
So here they're kind of just setting the scene with the basically how MCP servers are great blah blah blah but also how they consume so much context before you even sent your request to the LLM. So here they've introduced three new experimental beta updates. You got the tool search tool which allows claw to search tools to access thousands of tools without consuming its context window. You got programmatic tool calling as well which basically writes code to access those tools. And then you got tool use examples which demonstrates how to use tools again. So you don't have to expose the whole thing. And they kind of give a little breakdown of it here. You can see like it just slightly faster. Then by the end of it kind of like this tool the actual context window it's just so much less here. You can see it there.
Now here is the main event is the tool search tool which is a tool in of itself which uses reax to find other tools. And I sort of demonstrate here the context use just by being able to search tools rather than just provide them all in the context up front. Here we go. The tool search tool. Let's call dynamically search tools instead of loading all definitions up front. You provide all your tool definitions to the API but mark tools with defer loading true to make them discoverable on demand. Deferred tools aren't loaded into Claude Claude's context initially and you can still use defer loading false if you want specific tools available up front.
So here we go. Here's an example of the implementation here. Here is a quick and dirty example. I'm setting up my anthropic API key and a few other uh constants. You're going to be making a regular request to the anthropic API and you're defining advanced tool use as a flag in the request itself. If we go up to the tools in which I'm passing it, you'll see that we're including that um the tool search tool reax as the first tool and then I've got like a dummy get weather tool here. I've got a dummy get time tool here and I'm flagging these as defer loading. So none of these tool calls will get flagged uh will get written into the context up front. it will be the tool search tool that actually searches these when I make the request and I make the fetch request with the body along with those tools uh those headers as well and return in the response.
So you can still keep the traditional access to those tool calls that you do need up front and they highlight a few reasons why you would use it and when you wouldn't use it here. So tool definitions consuming over 10,000 tokens. I guess they consider under 10,000 tokens not very problematic. If you're experiencing tool selection um accuracy issues, building MCP MCP powered systems with multiple servers or you're using loads of servers basically.
Now the second solution here is obviously pro programmatic tool calling. And the long and short of it is that you can bundle tool calls together to again reduce the bloat that inevitably ends up back in the context. Again, it's just a simple line allowing that tool call to enter to allow to be called by the code execution agent and then Claude is able to actually write um Python code to to make those requests and then only the final output actually enters the context window.
And then finally, the tool use example. You're actually providing examples of the tool call in the JSON schema here for claw to be able to actually run. And once again, when it's beneficial, when it's not beneficial. This is really handy for for people building agents and things like that. This is such a great step forward in MCPS cuz there there is a huge amount of possibility for having MCPS available. It's just the context and and and making your AI dumber that I don't like.
So, what do you think? Are you as excited about these advancements as I am? Let me know down in the comments. Oh, patent disrupt. Also, I'm going to be getting more into agent stuff during the next few months. So, if you want to hear more about how to build agents and how to use stuff like this, then also let me know down in the comments. Or if you want early access, then head over to my Patreon, which is patreon.com/zerx5 for early access to those videos and advanced learning. Anyway, like, subscribe, and all the rest of it.