Alex

@sdrzn Looks like @AnthropicAI listened. I implemented this using the cline codebase many months ago. Also did GitHub based skills and a new customizabile chat UI/UX, check it out in the “Astro Code” VS Code extension x.com/x413x5/status/…

@sdrzn Also worked on redesigning the main UI/UX of the chat interface, the verbosity can be customized, to look and feel closer to what a user would prefer. Not released yet, working on merging your latest changes, and maybe we can integrate it back to @cline

It looks like subagent mcp config has been something ppl have been asking for since august, IMO makes a ton of sense! github.com/anthropics/cla…

Claude skills at first felt like dynamically loaded slash commands, but it’s so much more. They help avoid context rot by conditionally loading additional assets like case-dependent markdown instructions or scripts to execute for more deterministic outputs. But even that felt like overkill when slash commands could technically accomplish the same thing, until I realized how skills can be applied to subagents. Claude code can spawn subagents with specific skills to handle scoped work in a deterministic way, which really isn’t otherwise reliable due to the telephone game these agents play with each other. In other words, the best abstraction is using slash commands as a shortcut for YOU to instruct claude, and skills for this main claude to instruct subagents. It seems obvious then if we’re trying to reduce context rot that skills should be able to dynamically load specific mcp servers so that the main context window doesn’t get polluted with all of them. Curious why that isn’t the case (cc @bcherny @amorriscode )? This feels like a better approach than tool search tool since it prevents the wrong claudes from loading in tools they shouldn’t use. But I’m guessing this was a ux decision to strike a balance between convenience and complexity - since it would require users to set up skills with mcp’s to get the right behavior, whereas with tool search tool a strong model should do a good job of picking the right tools with zero config at the cost of another request and the lack of awareness of its own mcp capabilities, which might even lead to worse outcomes overall. (I imagine to combat this, users will give skill instructions about what mcp tools it can access via tool search and what they do, but this seems very duplicative.)

@sdrzn x.com/x413x5/status/…

@sdrzn I’ve been following and experimenting with the cline codebase. One simple solution that I did to avoid context bloat is to skip providing full MCP input schema, just tool name and descriptions. I’ve implemented a native tool that the agent calls to get the schema just in time.

@AnthropicAI A working example that greatly reduces upfront context token usage and loads on demand only the MCP tool definitions the agent is actually using (if any). An implementation based on @cline

@AnthropicAI Stop loading full MCP tool definitions upfront in the context. Only provide tool names and descriptions, and provide a native tool to the agent that it can call to load the full MCP tool definition just before actually using a new tool. This greatly reduces the token usage as is.

New on the Anthropic Engineering blog: tips on how to build more efficient agents that handle more tools while using fewer tokens. Code execution with the Model Context Protocol (MCP): anthropic.com/engineering/co…

@DeployStack This works great if you have your own agent client. As general approach your solution could be better, it depends on how many additional requests are needed to get to the final tool call.

@DeployStack Most of those tokens are usually for the tools input schemas. Another approach we implemented in our agents is to skip providing full schema for tools and, instead, instruct the agent to call a native tool that will provide the schema just in time before using a new tool.

Context window problem solved! DeployStack's hierarchical router cuts MCP token usage by 99.5%. Two meta-tools replace 150+ tools, freeing 81,650 tokens for actual work. Scale to 100+ MCP servers without performance limits. deploystack.io/changelog/hier… #MCP #AI #DevTools

Releasing the first standalone iteration of Astro, a customized version of @cline, as a foundation for domain specific coding agents that can have specialized functionalities and capabilities for different industries. It is available in the VS Code Marketplace as "Astro Code". Astro is the base for @xPilotAI, a domain specific coding agent for MultiversX.

Releasing the next iteration for @xPilotAI, featuring a GitHub expandable knowledge base, and more…

@sdrzn @cline @amandaksilver Congratulations to you and the incredible team behind cline ❤️ The cascading effect of being open source and investing in the agentic ai coding landscape is hard to quantify—but it’s clear it marked an inflection point in the history of how software is written.

@amandaksilver Congratulations to you and the incredible team behind vs code ❤️ The cascading effect of being open source and investing in the extension marketplace is hard to quantify—but it’s clear it marked an inflection point in the history of how software is written.

Happy birthday!