The LangWatch February Drop: Cheaper Events, Claude Code, and Multimodal Evals

February was a big month for shipping. We reworked how events are priced so building sophisticated agents does not become a punishment, wired LangWatch directly into Claude Code, taught our LLM-as-a-Judge to look at images, added conversation-level evaluation, and brought full templating logic to prompts.

Here is everything that landed this month.

A major cut to event pricing

Modern agents are getting more elaborate, and that has a real consequence: a single user interaction can fan out into dozens of internal steps, tool calls, and sub-operations. Teams should never have to think twice about building a more capable agent or running one more experiment just because it might inflate their bill.

So we rebalanced our pricing to reward that ambition instead of taxing it. The Growth plan now ships with 200k events included. Beyond that, extra usage runs $6 per 100k events. Collaboration stays simple at 29 euros per core seat, with unlimited lite seats so stakeholders can look in without added cost.

The result is spend you can actually forecast. Compare that to the typical LLMOps platform, where the invoice climbs steeply the moment your trace volume grows, and the difference in predictability becomes obvious as you scale.

Claude Code, the LangWatch MCP, and Scenario

Your coding assistant should be able to see your observability data, not just your source files. This month we made that possible: using the LangWatch CLI together with our MCP Server, you can plug Claude Code straight into LangWatch and give it direct access to traces, spans, performance metrics, and agent testing scenarios.

The star here is MCP Server v0.5.0, which lets Claude work with Scenario tests without leaving the editor. From natural language, it can list, create, update, and retrieve Scenario tests, effectively turning a plain-English edge case into a structured agent simulation. Developers can pull traces, track down the root cause of an issue, and refine their tests all inside the same workflow, no context switching required.

You can connect your workspace at app.langwatch.ai and learn more in the docs.

Multimodal evaluations

Plenty of AI systems produce output that only makes sense in light of an image, and until now our judge could not see that image. That changes this month: LLM-as-a-Judge now supports images, so the judge can factor the visual input into its verdict.

This is built for the workloads where vision matters, document processing pipelines, vision agents, and image-based assistants among them. When the judge evaluates an output, it now weighs both the generated text and the original image together, and it does this across entire workflows rather than a single isolated prompt.

Evaluation by thread

Sometimes the unit that matters is not one trace, it is the whole conversation. That is why we added thread-based evaluation, which lets you assess your LLM application at the conversation level.

With thread-based mapping, every time a trace gets evaluated, LangWatch pulls in the full context of the conversation thread it belongs to. That means you can study how an agent behaves across a complete interaction, and quickly spot which threads are performing well and which are struggling.

LiquidJS support

Static prompt templates only take you so far. This month we added full LiquidJS support, giving you complete logic control inside your prompt templates. You can now loop over lists, branch with conditionals, and assemble prompts dynamically instead of settling for a single fixed variant.

There is a migration bonus, too: because the same Liquid-style templating carries over, moving your prompt templates from Langfuse to LangWatch is far less painful, since you can reuse what you already wrote.

Looking ahead

There is more coming. We are pushing toward deeper agent simulations and richer quality signals, and we are opening a first beta to simulate voice agents. The last spots for that beta are filling up quickly.

We also have a few webinars on the way: Defining Agent Quality, Automating Evals with Claude Code, and AWS Bedrock plus LangWatch monitoring agents.

LangWatch is the open-source LLM evaluation and agent testing platform.