I've been really impressed with the LiveKit MCP server. I had a stubborn LiveKit connection bug that my agents and I were spinning on for a while. After chewing through a lot of tokens, I started a fresh Codex session with the LiveKit MCP server enabled and, within minutes, the agent zeroed in on the root cause and the fix. (I use Claude too, but this particular run was with Codex.) That success immediately raised a practical question: Should I leave the LiveKit MCP server enabled all the time? The naive answer is "sure, why not?" But there's a hidden cost: tool definitions take up context, and in many agentic setups those tool definitions are effectively included on every model inference call. If that's true for your environment, the overhead is paid even when the agent never calls a LiveKit tool. This isn't unique to LiveKit — it applies to any MCP server with a sizable tool catalog. LiveKit just happens to be the one I measured. My rough token overhead estimate To get a back-of-the-napkin number, I wrote a script to pull the MCP tool metadata (schemas + descriptions) from the LiveKit docs MCP endpoint (

<https://docs.livekit.io/mcp>

) and used that as a proxy for what gets injected into the model context when the MCP server is "loaded." Based on the size/structure of that tool catalog, I estimate the LiveKit docs MCP server adds on the order of ~3k input tokens per model inference call (ballpark range: ~2.8k–3.5k), even if none of the tools are ever used. A caveat on the estimate: the actual token count depends on how your agent framework serializes tool definitions (JSON schema, markdown, etc.), so your mileage may vary depending on the framework. Now, that doesn't sound awful… until you multiply it by how many inference calls a typical coding session makes. What does that mean over a whole session? As a concrete example: if an agentic coding session triggers ~180 model inference calls (very plausible if it iterates, runs commands, reads outputs, revises, etc.), then the unused overhead would be: • ~3,100 tokens/call × 180 calls ≈ 558,000 extra input tokens • …just to carry tool definitions that never get invoked. To be clear: that 558k number is not universal. It's an example that depends entirely on (a) whether your agent resends tools each inference call, and (b) how many inference calls your session typically makes. Cost Estimates Metric Per Inference Call Full Session (180 calls) LiveKit MCP Overhead ~3,100 tokens ~558,000 tokens Estimated Cost (GPT-4o) ~$0.007 ~$1.39 Estimated Cost (Claude Sonnet 4) ~$0.009 ~$1.67 (Costs are rough estimates based on published input token pricing. Verify against current rates for whatever model you're using.) My rule of thumb now I start sessions with zero MCP servers enabled and add them only when I know I'm going to use them. In practice that means: edit the MCP config, reload, and continue the session once the tools are actually relevant. Notes / caveats These are rough estimates and will vary by model, tool encoding, and agent implementation. • If your environment supports prompt caching (like Claude's ephemeral caching), this "tax" is mostly a latency/first-call hit. But for standard API usage, it's a pure recurring cost. • If anyone has hard numbers from logs or a token counter for their setup, I'd love to compare notes. Bottom line: the LiveKit MCP server is genuinely useful, but I'd treat it like a power tool — keep it nearby, not running continuously.

Thank you for this feedback, this is a really interesting breakdown that you've done a great job of walking through. I'm going to go ahead and forward this to the MCP server developers