LangWatch is an open-source (Apache 2.0) LLM engineering platform for teams running AI in production. You trace every LLM call with OpenTelemetry. You test agents against simulated text and voice users, score outputs with LLM judges or custom code, and route model calls through one gateway with budgets. It also tracks coding agents like Claude Code. Use the managed cloud, self-host it, or run a hybrid setup. It fits engineering, product, and QA teams shipping agents.

• Agent Simulations: Run simulated text and voice users against your agent over many turns. You can write scenarios in plain language, run them locally or in CI/CD, and use them for red teaming. A judge agent reads the full trace and explains each verdict.
• Evaluations: Score single outputs or whole conversations with an LLM judge, custom code, or a full workflow. Run evals from a notebook or the UI, on live production traffic, or pairwise to compare models, prompts, and versions.
• OpenTelemetry-Native Observability: See token, cost, and cache accounting for every span, with waterfall, flame graph, topology, and sequence views. Plain-language search, saved views, and automatic topic clustering help you find problems fast.
• AI Gateway: Send calls to many model providers through one endpoint using your own keys. You can set budgets per person, team, or customer, and LangWatch cites about 11 microseconds of overhead.
• Prompt Management: Version, deploy, and A/B test prompts as code, with full history and GitHub sync.
• Langy AI Engineer: An assistant that reads traces, writes tests, and opens pull requests. It turns a PM's plain-English goal into a test plan and turns regressions into draft prompt fixes for developers to review.
• Coding Agent Tracking: Track token use, cost, and history for Claude Code, Codex, Cursor, GitHub Copilot, and other coding agents with one command.
• Flexible Deployment: Pick managed cloud in EU, US, UK, or APAC regions, self-host with Docker or Kubernetes/Helm, or run a hybrid setup where your data stays on your own infrastructure.

• Open source and self-hostable: The Apache 2.0 license and self-hosting on any plan give you control over your data and your infrastructure.
• Wide framework coverage: You get Python, TypeScript, and Go SDKs, plus support for LangChain, LangGraph, CrewAI, Mastra, Pydantic AI, Vercel AI SDK, OpenAI Agents, DSPy, and plain OpenTelemetry.
• Simulation-based agent testing: Multi-turn simulated users and judge agents let you catch regressions before you deploy, which goes further than single-response evals.
• Usable free tier: The Developer plan costs nothing and needs no credit card. It includes 50k events a month, so you can try it on a real project.
• Product teams can join in: Plain-language scenarios, AI search, and Langy let PMs and QA help with testing without writing all the code themselves.
• Enterprise compliance: ISO 27001 certification, GDPR compliance, EU data residency, and SSO/SCIM support help it pass security reviews at regulated companies.

• Tight free-tier limits: The Developer plan allows only 2 users, 14 days of data access, and 3 scenarios, 3 simulations, and 3 custom evals. Teams that want ongoing testing will need to upgrade quickly.
• Governance features are Enterprise-only: Custom SSO and RBAC, SCIM provisioning, audit logs, data masking, and hybrid or on-prem deployment are on the custom-priced Enterprise plan. Mid-sized teams with security requirements will need to talk to sales.
• Self-hosting takes infrastructure work: The self-hosted version runs on Docker Compose with your own ClickHouse. You take on database operations and upgrades that the managed cloud would handle for you.
• Built for engineering teams: LangWatch is designed for teams building agents and LLM pipelines. If your app only makes simple single LLM calls, a lighter logging tool may cover what you need for less setup.

• Pre-release agent regression testing: A team building a customer support agent writes scenarios for refunds, edge cases, and adversarial prompts, then runs them in CI/CD on every pull request. Broken behavior fails the build before it reaches customers.
• Production quality monitoring: You run online evaluations on live traffic and cluster conversations by topic. Drops in quality after a deploy show up on dashboards, and you can trace them to the exact span.
• Voice agent QA: Teams shipping voice assistants simulate callers over many turns to test how the agent handles interruptions, tool calls, and unclear requests, without manual test calls.
• Model and prompt comparison: Before switching providers or rewriting a prompt, you run pairwise evals on the same dataset. You see which version wins on quality and cost before you roll it out.
• LLM cost control across teams: Platform teams route all model calls through the AI Gateway and set budgets per team or customer. Finance gets per-span cost attribution instead of one big provider bill.
• Coding assistant usage tracking: Engineering leaders connect Claude Code, Codex, or Cursor to see token spend and usage history across developers.