FrontMCP adds hot reload for MCP servers: 0.5s restarts, no Claude Code reconnects

DavidAntoon · reddit · 2026-10-27

The author of FrontMCP, an open-source TypeScript MCP server framework, demonstrates eliminating the "edit → restart → /mcp reconnect" loop when building MCP servers with Claude Code.

Setup (three commands):

npx frontmcp create notes-server

cd notes-server && npm install

claude mcp add --scope project notes-server -- npx frontmcp dev --stdio --serve

The last command writes .mcp.json so the whole team gets the same server.

Measured results (Claude Code 2.1.310, Opus 5.5): asking Claude to call add, add a multiply tool and call it, then make add reject negative numbers — every save restarted the server in 0.44–0.62s, calls returned the new behavior immediately, and the whole flow took 58 seconds with zero manual restarts or reconnects. Control test: same prompt over HTTP to npm run dev (tsx --watch) took 6s per restart plus one ECONNREFUSED retry.

How it works: the stdio bridge process Claude Code starts never exits; it runs your server as a child over a pipe, watches files, restarts the child on change, holds in-flight calls, and sends notifications/tools/listchanged (0.6s after save) so Claude Code refreshes the tool list without a reconnect. If a save breaks the server, calls wait up to 30s, then fail with -32097 devreloaddeadline and the real error lands in .frontmcp/dev.log — the post suggests pointing Claude there via a CLAUDE.md note.

Original post →

More from coding & agent

coding & agent channel →