Buildkite developed an MCP server that helps AI agents analyze CI/CD build logs more effectively. The team discovered that raw terminal streams with ANSI codes overwhelmed LLMs, causing them to fixate on early errors rather than root causes. They solved this by preprocessing logs into Parquet format with structured entries (timestamp, content, group, flags) and creating four navigation tools: tail_logs, search_logs, read_logs, and get_logs_info. Using Claude to critique its own debugging attempts, they iteratively refined the toolset to mirror human debugging workflows, making correct analysis paths the easiest for agents to follow.
Table of contents
Background on our MCP serverWhy CI logs are a challenge for AI agentsOur first attempt: a tail toolDesigning log-navigation toolsWhat we learnedSort: