Hi all Livekit team and community :wave: Curious ...
# agents
b
Hi all Livekit team and community 👋 Curious how people are building around Tool calling and MCPs. Specifically about giving agents a dedicated and isolated execution environment to let them write code that calls MCPs or Tools via a generated API. In contrast to the traditional MCP/Tool definition passed to the LLM context window and letting the LLM chose. These two articles explain it very well: • https://blog.cloudflare.com/code-mode/ • https://www.anthropic.com/engineering/code-execution-with-mcp The main benefits on this approach: • Huge (90%+ in some cases according to Anthopic) tokens saved. • Less context bloating for llms • LLMs hallucinate less due to shorter contexts and being much more versed at writing/reading code than MCP definitions. I have 2 questions: • Is Livekit Tool implementation more like traditional MCP? (my understanding is yes) • Has anyone in the community started implementing tool calling like code execution?
r
For question one yes, LiveKit agents pass MCP servers and function tools directly to the LLM context, so traditional. I don't know what others are doing in the community.
✅ 1
b
Thanks @refined-appointment-81829! What's your take on tools-as-code for LLMs? Have you given a look? Reducing the tokens on the context window can save much LLM processing speed, but at the same time, running code somewhere else adds an additional component of network latency. I wonder where the breakeven point is.