The amazon-bedrock-agentcore-mcp-server crashes on startup and never completes MCP initialization because main() calls cache.ensure_ready(), which fetches the hardcoded docs index at https://aws.github.io/bedrock-agentcore-starter-toolkit/llms.txt. That URL now returns HTTP 404 (the upstream docs site is deprecated and the llmstxt plugin was removed), and the fetch has no error handling, so the urllib HTTPError escapes main() and the process exits. Every tool group registers fine; only the docs index fetch kills it.

This wraps the per-source llms.txt fetch in load_links_only() in a try/except that logs a warning and continues, and still marks the index loaded so ensure_ready() does not retry on every call. The documentation search tools degrade to returning empty results while the runtime, memory, identity, gateway, policy, browser, and code interpreter tools keep working.

I also fixed test_docs_tool_invocation, whose len(result.content) > 0 assertion only passed because the 404 turned the call into an error response with text content; it now checks the tool stays callable and returns a well-formed (possibly empty) result set. Added a unit test for the graceful-skip path. Full server test suite (813 tests) passes and ruff is clean.

Fixes #4138

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of the project license.