MCP tool names advertised by server don't match names exposed to model

Summary

MCP servers advertise their tools using their own naming convention (e.g. agentbridge-read_file), but Claude Code exposes those same tools to the model under a transformed name (mcp__agentbridge__read_file). The MCP server's system-reminder instructions use the server's own naming convention when describing which tools to call. The model sees both the instructions (with agentbridge- prefixed names) and the actual callable tools (with mcp__agentbridge__ prefixed names) but cannot reconcile the two, leading to tool-not-found errors that the model misinterprets as connectivity failures.

Expected behavior

Either:

• The system-reminder tool name references should be rewritten to the mcp__servername__toolname form before being injected into context, so the names match what the model can actually call, or
• The model should be explicitly informed of the name mapping between the server-advertised names and the Claude Code-exposed names.

Environment

• Client: Claude Code