Tools
Langfuse 4.27.0 adds an MCP evaluator testing tool and multimodal eval inputs, plus cached input token cost in the trace table
Published 2026-09-01, v4.27.0 adds an evaluator testing tool to Langfuse's MCP server (#16870) so evaluators can be exercised from a coding agent, supports multimodal inputs in evals (#16675), and surfaces cached input tokens and cost as trace-table columns (#16767). Normalized trace I/O parsing is now hooked up behind a feature flag (#16775), continuing the normalized I/O contract work started in v4.25.0. The prior day's v4.26.0 added linking from a trace to the evaluator execution trace that scored it (#16839), which closes the loop on debugging why an eval produced a given score.
Source
↳ Follow the thread