MCP client setup¶
Start the Prism MCP server, then point your client (Claude Code, Claude Desktop, Cursor, VS Code Copilot, …) at it.
Start Prism¶
Leave the terminal open — the server runs in the foreground. Stop with
Ctrl+C.
The default URL is http://localhost:3000/mcp. Every client example
below assumes that URL; adjust the host/port if you ran
prism serve --port 4000.
Clients¶
Edit .claude/settings.json (in your project, or in %USERPROFILE%
for a global config on Windows):
Claude Code connects over HTTP to the running server — you start Prism manually, Claude Code doesn't launch it.
Edit claude_desktop_config.json, located at
%APPDATA%\Claude\claude_desktop_config.json:
Claude Desktop will launch prism serve as a child process when it
starts, and stop it when it exits. Works because prism.exe is on
the system PATH after the installer.
Edit .cursor/mcp.json (workspace) or ~/.cursor/mcp.json (global):
Same pattern as Claude Code — start prism serve yourself.
Enabling the interactive debugger¶
Debug tools are registered only when IRIS_DEBUG_ENABLED=true. Start
Prism with that env var set if you want Claude (or any other client) to
have access to debug_start, debug_step, etc.:
See Interactive debugger for the full workflow.
Enabling workspace mode¶
Workspace mode unlocks two additional MCP tools (put_document,
put_and_compile) that read files from a local directory. Set
IRIS_WORKSPACE to the path you want the AI to push code from:
Without it, the tools are not registered (and the startup log shows
workspace: off).
Verifying the connection¶
Ask your client to call get_server_info — a zero-argument tool that
returns the IRIS version. Any response means the transport is wired up
correctly. If the AI says the tool isn't available, re-check the client
config file's path and JSON validity.