MCP server consumes 24k tokens just by connecting - makes it unusable

[Luis Rudge]

For reference, asking "1+1?" with Jira MCP enabled costs 24,488 tokens. The same query without it costs a few hundred.

This is a significant problem because:

• Most LLM context windows are 128k-200k tokens - Jira alone eats 12-20% just by existing
• Users pay per token - this adds unnecessary cost to every single request
• It makes the MCP impractical for real coding workflows where context is precious

The tool definitions being sent are clearly too verbose. Please consider trimming the schema descriptions or implementing lazy-loading for tool definitions.

Please implement dynamic tool discovery: https://modelcontextprotocol.io/legacy/concepts/tools#tool-discovery-and-updates

[Ibby]

Thanks Luis for the feedback, we will work on addressing this issue.

[Luis Rudge]

Thanks!

[Ibby]

Hey @Luis Rudge 

We started rolling out tool level description reduction for both Jira and Confluence, it would take until end of the week to fully rollout to all regions. The reduction for initial connection is ~50% of tokens from tool description and metadata.

Next we will start working on improving the responses like search tools etc to avoid returning unused fields.

[Luis Rudge]

Thank you! Excited to give it a shot when it's live

[Ibby]

The change is now live on 100%.

We will start with reducing response token size next.