Langwatch

LangWatch is an Amsterdam-based company that builds an open-source platform for testing, evaluating and monitoring LLM-powered applications and AI agents. The product spans four main areas: agent simulations, in which a simulated user drives multi-turn text or voice conversations against a customer's agent; evaluations, which score single outputs or entire conversations using LLM-as-a-judge, custom code or workflows, offline and on live production traffic; observability, built on OpenTelemetry with support for the GenAI semantic conventions, covering traces, spans, tokens, cost and cache accounting, topic clustering and custom analytics; and prompt management, including a prompt registry, prompt playground and a GitHub integration that keeps prompts in Git and links prompt versions to traces.

The platform also ships an AI Gateway — an OpenAI/Anthropic-compatible proxy distributed as a separate Go binary and Helm sub-chart — providing virtual keys, hierarchical budgets, inline guardrails and automatic fallback across providers, plus AI governance features. Adjacent open-source components include Scenario, the simulation framework, and a Better Agents CLI, along with MCP servers and "Skills" that let coding assistants such as Claude Code write and run agent tests. Red-teaming is offered as adversarial simulation that probes agents for jailbreaks, policy violations and unsafe tool calls before release. Agents can be tested in whitebox or blackbox mode, through the API or by hooking into internals, and tool calls, skills and MCP servers are traced and mockable for deterministic runs.

LangWatch is deployable as managed multi-tenant SaaS (EU, US, UK and APAC regions), fully self-hosted via Docker Compose, Kubernetes/Helm or a customer VPC, or in a hybrid model with the data plane on customer infrastructure and the control plane hosted by LangWatch. Enterprise controls include RBAC and REST APIs, SCIM and SSO, cost-center attribution, audit log export to SIEM and custom retention policies; the company states ISO 27001 certification, GDPR compliance, EU data residency and monitoring via Vanta.

Founding story

Co-founders Manouk Draisma and Rogério Chaves met at Antler Amsterdam, where the GenAI side projects each was building turned out to share the same missing piece: quality control. Draisma had come from commercial leadership roles at Lightspeed and Biller; Chaves had spent years at Booking.com, where he helped build its first AI agents and saw how fragile and hard to test such systems become at scale, following an engineering background at ThoughtWorks. Their shared experience was that it was very difficult to understand how users interacted with GenAI systems and to communicate the resulting issues to engineers, which led them to build the evaluation and observation platform they wanted for themselves.

Business model

Commercial open-source: the platform source is available and can be self-hosted for free, while LangWatch monetises a hosted cloud service and enterprise deployments (hybrid and on-premises) with security, governance and collaboration features. A free cloud plan is offered for self-service sign-up, with demos and solutions-engineering onboarding for larger accounts.

Subscription access to LangWatch Cloud, starting from a free plan, plus paid enterprise/self-hosted and hybrid deployments; pricing was revised in February 2026 with the stated aim of decoupling customer bills from growth in AI usage.

Traction

The open-source repository langwatch/langwatch has approximately 3,500 stars, 355 forks and over 6,500 commits. The company publishes named customer testimonials from founders, AI engineers and technology leaders describing use in production, including for payment-related interactions, and maintains a customer stories section. At the time of the 2025 pre-seed announcement the company said early customers were able to test, iterate and ship more than ten times faster using the platform.

Latest developments

Recent releases centre on Langy, an assistant that converts a plain-language product goal into a Scenario test plan, selects the simulator, generates scenarios and a JudgeAgent rubric, runs simulations in parallel, scores them and drafts prompt revisions as pull requests shipped through the Prompt Registry; the company cites a median goal-to-pull-request time of 14 minutes. Other 2026 developments include Claude Code usage and token-cost tracking, multimodal evaluations, voice-agent testing, LangWatch MCP and Skills for coding assistants, v3.0, a new pricing model, and the release of the entire platform as open source.