LangWatch Review (2026): Pricing, Features & Honest Verdict

LangWatch is an LLM engineering platform that tests, evaluates and monitors AI agents before and after they reach production. It covers more ground than the observability tools it competes with, the whole platform is Apache 2 licensed, and you can self-host it for free. The per-seat cloud pricing gets expensive for large teams. Best for: engineering teams shipping agents they cannot afford to have misbehave. Price: Free forever, then from EUR 29 per core-seat/mo (free plan: yes). Rating: 8.5/10.

What is LangWatch?

LangWatch is a platform for making AI agents reliable. The pitch is that agents are unpredictable and most teams ship them anyway, having tested roughly nothing, because the tooling for testing non-deterministic software barely existed until recently. LangWatch bundles the pieces that address it: simulations that run realistic user scenarios against your agent, evaluations that score response quality, observability that traces every step and tracks cost and latency, and governance that controls which models and keys get used at all.

The company is LangWatch B.V., built in Amsterdam. It holds ISO 27001 and publishes GDPR documentation, which matters because a tool ingesting your production traces handles some of the most sensitive data you have. The platform itself is open source under Apache 2, with the main repository sitting at around 3,500 stars, and the Scenario testing framework is a separate Apache 2 project with Python and TypeScript SDKs. We found LangWatch through its Claude Code usage tracking launch, which turned out to be one feature on a considerably larger platform.

What Are LangWatch’s Key Features?

Scenario testing and simulation

This is the part we would actually pay for. LangWatch runs realistic user scenarios against your agent so you catch failures before production rather than in a support ticket. Testing a deterministic function is easy. Testing an agent that can take twelve different paths through a conversation and call six tools along the way is not, and it is the reason most teams ship agents on vibes. The Scenario SDK being open source means your test definitions are portable and you can run them in your own CI without asking permission or holding an account.

Evaluation and observability in one place

Evals measure whether responses are actually any good, and observability traces every agent step while monitoring cost and latency in production. Most teams currently buy these separately or build half of each. Having them in the same product means the traces that show a regression can feed directly into the eval set that catches it next time, which is the loop that makes the category worth anything. The platform also does prompt management, so prompts are versioned, deployed and A/B tested as code with full history and GitHub sync rather than living in a spreadsheet.

An AI gateway with real governance

LangWatch will sit in front of your model calls and govern them. You get virtual keys with budgets attached, routing policies that decide which model serves which request, and a full audit trail. For anyone who has watched an experimental agent quietly spend four figures over a weekend, budget-capped virtual keys are the feature that pays for the subscription. It also solves a procurement problem, because it gives you one place to answer which models and tools are in use across the org.

Voice testing, red-teaming and coding agent tracking

Three narrower capabilities round it out and each targets a real gap. Voice AI testing simulates calls against voice agents at scale before they speak to a customer. LLM red-teaming runs simulated attacks to surface safety and security holes. And the Claude Code usage tracking that brought us here gives full trace history and token spend for Claude Code, Codex and other coding agents, which is a genuinely useful thing to have when your engineering team’s AI bill is growing and nobody can say why. LangWatch also advertises an automated AI engineer that reads your traces, writes the tests and opens the pull request.

How Much Does LangWatch Cost?

LangWatch is free to start and then charges EUR 29 per core-seat per month. The Developer plan is free forever with no credit card: 50,000 events a month, 14 day data access, 2 users, and 3 scenarios, 3 simulations and 3 custom evals, with community support through GitHub and Discord. That is a real free tier rather than a trial, and for a solo developer or a side project it may be all you ever need.

Growth is EUR 29 per core-seat per month and adds 200,000 events included with overage at EUR 5 per 100,000, 30 day retention with extension at EUR 3 per GB, unlimited lite-users, unlimited simulations, evals and prompts, and private Slack or Teams support. Volume discounts kick in above 20 users. The core-seat versus lite-user split is the pricing decision that makes this work: you only pay full price for the engineers building and testing, while everyone who just needs to look at dashboards is free. A lot of competitors charge per seat flat and punish you for giving product managers access.

Enterprise is custom and covers hybrid or on-premise deployment, custom retention, SSO and RBAC, audit logs and SLAs, ISO 27001 reports, custom terms and a DPA, a forward deployed engineer, and billing through the AWS or Google marketplaces. Against the field this lands well. Langfuse runs a free Hobby plan then $29, $199 and $2,499 tiers with a $300 a month teams add-on. LangSmith is free for a developer seat then $39 per seat per month. Braintrust starts free with $10 of credits and jumps to $249 a month for Pro. LangWatch’s EUR 29 core-seat sits at the cheap end while covering more ground than any of them.

What Does the Open Source Licence Actually Get You?

More than usual, and this is the detail that moved our rating. A lot of vendors in this category describe themselves as open source when they mean a client SDK is on GitHub and the product is not. LangWatch is not doing that. The main platform repository is Apache 2 licensed, and the company’s own FAQ states that LangWatch runs fully self-hosted with docker compose against your own ClickHouse instance, so nothing leaves your environment. Enterprise self-host is a paid upgrade, but what it adds is SSO, RBAC, SLAs and support rather than the ability to run the software at all.

That distinction is worth money. It means the free path is not merely a limited cloud tier, it is the whole platform running on your own infrastructure with no event ceiling and no retention window imposed by a plan. If your objection to hosted observability is that you cannot send production traces to a third party, that objection does not apply here. Apache 2 is permissive rather than copyleft, so there is no obligation to open your own code, and there is no clause reserving commercial use.

The trade you are making is operational rather than legal. Self-hosting means you run and upgrade ClickHouse, and you own the reliability of a system whose whole purpose is telling you when something else broke. For most teams under twenty engineers the cloud plan is the better use of time. But the option being genuinely free changes the negotiating position, and it means a decision to pay LangWatch is a decision about convenience rather than access.

Who is LangWatch Best For?

Use LangWatch if you have agents in production or heading there and you currently have no systematic way to know when they degrade. It is the strongest fit for teams of roughly three to twenty engineers, where the core-seat model is cheap and you have enough surface area to need simulation rather than spot checks. It is also the obvious pick if you are in a regulated or European context, given the ISO 27001 posture, GDPR documentation and the fact that a fully self-hosted deployment is available without a contract. And if your problem is specifically that nobody knows what your coding agents are spending, the usage tracking solves that on the free tier.

Skip LangWatch if you are making simple single-turn model calls with no agentic behaviour, because tracing and evals are overkill and a logging library will do. Skip the paid cloud plan if you have a very large team of full contributors, since EUR 29 per core-seat multiplies faster than a usage-based competitor would, and at that size self-hosting starts to look sensible. And skip it if you want a managed service with no operational choices to make, because the flexibility here comes with decisions attached.

Best LangWatch Alternatives

Langfuse is the closest open source competitor and the most direct comparison, since both offer a permissively licensed platform you can self-host. Pricing runs from a free Hobby plan to Core at $29, Pro at $199 and Enterprise at $2,499, with a teams add-on at $300 a month. Langfuse is the more mature tracing product with a larger community. LangWatch is stronger on simulation, red-teaming and gateway governance, which Langfuse does not really attempt.

LangSmith from the LangChain team is free for a developer seat and $39 per seat per month on Plus, with usage-based charges on top and a custom Enterprise tier. It is the default if your stack is already LangChain or LangGraph, and the integration advantage is real. It is proprietary, it has no self-host path outside enterprise, and it has nothing equivalent to the AI gateway.

Braintrust is the eval-first option, starting free with $10 of credits and moving to $249 a month for Pro. It is excellent at evaluation workflow and dataset management, and considerably more expensive than LangWatch for a small team. Choose it if evals are the whole job and observability is someone else’s problem.

Final Verdict: Is LangWatch Worth It?

LangWatch is doing the unglamorous work the agent boom skipped. Everyone shipped agents. Almost nobody built a way to know whether those agents still work this week. LangWatch covers testing, evaluation, observability, prompt management and gateway governance in one product, at a price that undercuts tools doing a fraction of it, with a free tier that is genuinely usable. The Amsterdam base and ISO 27001 certification make it an easy internal sell in Europe.

What pushes it above the rest of the category is that the open source claim survives inspection. The platform is Apache 2, it self-hosts with docker compose on your own ClickHouse, and the paid tiers sell convenience, support and compliance rather than access to the software. That is the arrangement buyers should want and it is rarer than the marketing in this space suggests. The honest caveats are that per-seat billing scales awkwardly past twenty engineers and that self-hosting hands you an operational burden. Neither is a reason to pass. Start on the free Developer plan, wire up the Scenario SDK, and you will know within a week whether the simulations catch things your current process does not. We think they will.