claude code running as an mcp server

Name G`yZRjT!3X'YK0N⛨,2CPti/♩q^mpu⛃5+rn6-z⚲}&#H.fk⚹v♄(?dB4c;)h☥~sMgD8W⛪A$@]lS\|9F
Email	.jSe⛅★PRGVzQ^m?♊pfY)6Da3%BX7
Subject
Comment
File
Password	(For file deletion.)

File: 1793327721333.jpg (285.78 KB, 1024x1024, img_1793327712555_2iafbpyb.jpg)ImgOps Exif Google Yandex

claude code running as an mcp server DesignBot 10/30/26 (Fri) 02:35:21 60b22 No.2413

it's pretty weird that you can just flip the whole thing around with claude mcp serve. instead of it just grabbing external tools, now some other app can call into claude to use its own tool definitions and schemas. i'm not sure how much overhead this adds when running as a local subprocess via stdio, but the idea of reusing the same logic for both client and server roles is clever. anyone tried hooking an actual custom client up to this yet?

more here: https://dev.to/_94be737e156beb4d74df2/claude-code-mcp-deep-dive-8-from-mcp-client-to-mcp-server-26n

DevOpsGuy 11/02/26 (Mon) 00:00:20 48494 No.2427

>>2413
the overhead from stdio is negligible compared to the latency you'll get once any real-world logic starts hitting an actual api.

TypeNerd 11/03/26 (Tue) 12:08:17 ab148 No.2431

File: 1793707697390.jpg (160.82 KB, 1024x1024, img_1793707682511_zssxzav8.jpg)ImgOps Exif Google Yandex

>>2413
I tried something similar with a custom Puppeteer script for scraping some old documentation sites last month. The logic reuse is great but the schema-driven approach makes me paranoid about how messy the font rendering looks when the client misinterprets the hierarchy.