Scripts vs MCP
retained text
Working more with elaborate workflow skills that fetch infos from different apis (slack, gitlab, jira), i noticed limitations of MCP connections compared to CLIs. For example: i want to get all slack messages and their threads from the last 30 days in a slack channel, but without messages that already have a ✅ emoji response. Since i’m using the slack MCP, my agent has to first fetch all slack messages from the last 30 days, then filter out the ones with ✅ and fetch the relevant threads. That are multiple llm requests and multiple tool calls, and also all slack messages with ✅ land in the context unnecessarily. I at least instruct it to use a subagent so the impact on my main context is smaller. It’s just wasteful. With a slack CLI i could just use a single reusable bash script that makes multiple slack requests to only fetch the relevant threads, which costs only a single bash tool call and keeps my context clean and only requires a single llm request and no subagent. I miss this flexibility with MCP. If the MCP server does not already provide a tool that does exactly what i need, it’s going to be wasteful. It would be interesting if MCP tool calls could be composed script-like in a single tool call, probably even across multiple different MCP servers. What do you think about this? Do you also see it as a problem? Any suggestions how to solve it, potentially by extending the MCP standard? Should i create a github issue for the MCP repo?
document fetched 2026-10-25 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1xdyqwx/scripts_vs_mcp