LangWatch v3.0: April 2026 Product Updates

April turned out to be the heaviest release month in LangWatch history. The headline is a major version bump, but the story underneath it is bigger: a platform that now stands up an entire evaluation and agent-testing workflow in the time it used to take to read the setup guide.

Here is everything that landed.

LangWatch v3.0 is here

The centerpiece of the month is LangWatch v3.0. At its core sits a brand new trace storage backend that was engineered from the ground up for scale. Queries return faster, and self-hosting is dramatically less of a chore than it used to be.

To make that last point real, Helm chart support is now production ready, so running LangWatch on your own cluster is a supported, first-class path rather than an experiment. We also rewrote the documentation from top to bottom so that the self-hosting docs actually match how the product behaves today.

Agent Skills: an eval platform in a single session

The way people start with LangWatch has fundamentally changed. Roughly nine out of ten new users now begin through Agent Skills rather than clicking through the UI.

The idea is simple. You install the LangWatch MCP once, and your coding agent, whether that is Claude, Cursor, or Codex, immediately gains six purpose-built skills. From there the gap between "I want evals" and a running evaluation suite collapses from sprints down to minutes. The same holds for spinning up agent simulations.

Migrations get the same treatment. If you are coming from Langfuse or LangSmith, Claude reads whatever instrumentation you already have in place and rewrites it for LangWatch, so switching is no longer a project you have to schedule.

Red teaming: Scenario now thinks like an attacker

We publicly launched automated red teaming, and alongside it we relicensed the entire Scenario framework under Apache 2.0.

Here is the reasoning. A single adversarial prompt is not a real security test. Genuine attackers do not lead with the dangerous ask. They build rapport across many turns, then slip in a request that any model would have refused cold on turn one. Scenario now automates exactly that multi-turn pattern, so your defenses get pressure-tested the way they would be in the wild. See the red teaming docs to try it.

Goodbye Next.js, hello Vite

We migrated the application off Next.js. The app never relied on server-side rendering, so Next.js was carrying weight we did not need. The frontend now runs on Vite, and the API layer runs on Hono.

The payoff is tangible across the board: builds are dramatically faster, moving around inside the app is close to instant, and our CI integration test suite went from about an hour and a half down to nine minutes.

Enterprise: SCIM 2.0, Lite Members, and a new RBAC model

Three enterprise upgrades shipped together.

First, SCIM 2.0 provisioning. You can now provision and deprovision users automatically from Okta, Azure AD, or any SCIM-compliant identity provider through endpoints that follow RFC 7644. We also added an Auth0 log stream integration and an admin interface for managing scoped bearer tokens.

Second, Lite Members. This is a lower-cost user type built for the people who read traces and weigh in on evaluations without needing to touch infrastructure, think stakeholders, product managers, and domain experts.

Third, a full RBAC overhaul. Access is now built on role bindings, so a person can be a viewer on team A and a contributor on team B at the same time. It supports group-based access and permissions that are scoped to individual projects.

The new Traces UI (Beta)

We rebuilt the Traces experience specifically for how agents actually work today: multi-turn, multi-tool, and multi-agent. A few highlights:

Density modes. From a single saved lens you can flip between a data-dense view for engineers and a more readable view for PMs and domain experts.
Live data from the browser. Seed traces directly from your browser using a personal access token.
Ask AI. A built-in assistant helps you dig into errors and pick apart individual traces.

Looking ahead

The next product on the horizon is the AI Gateway, currently in private beta.

The premise: every app, SDK, or coding CLI routes its LLM traffic through one endpoint that is compatible with both OpenAI and Anthropic APIs. On every single request, that endpoint applies policy, enforces budgets, attributes cost, caches, runs guardrails, and handles fallback.

What that buys you:

Virtual keys scoped per team, project, or engineer, each with hard budget limits.
Real-time cost attribution so you always know where spend is going.
Automatic fallback chains, for example OpenAI to Anthropic to Azure to Bedrock to Vertex to Gemini.
Drop-in compatibility with Claude Code, Codex, Cursor, Aider, and Gemini CLI, no changes required.
Negligible overhead. Written in Go, it adds roughly 11 microseconds gateway-side at 5,000 requests per second.

Early access is open, so book a call if you want in.

You can dig into any of this at app.langwatch.ai or in the docs.

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

LangWatch v3.0 and the April 2026 Product Drop