Deploy and Host LangWatch (AI Observability) on Railway

Self-host the LangWatch application and background workers with NLP and evaluation services, PostgreSQL, Redis and ClickHouse.

About Hosting LangWatch (AI Observability)

The saved template contains seven container services: app, workers, nlp, langevals, Postgres, Redis and clickhouse. PostgreSQL, Redis and ClickHouse have persistent volumes. These are deployed containers whose upgrades, backups and access you manage; the template does not include a separate fully managed database subscription.

First deployment

Deploy and allow application initialization and database migrations to finish. Open the generated HTTPS domain on app.
Use the application’s email/password authentication flow to create your user and organization/project. The saved setting is NEXTAUTH_PROVIDER=email; no shared administrator account is supplied.
Obtain a project API key in LangWatch, configure your application’s LangWatch integration with that key and this self-hosted endpoint, then send and inspect a sample trace.
Add provider credentials for any evaluations or other integrations you choose to use. Configure outbound email separately if you need mail features.

Already configured

Generated app secrets, worker references, database connections and volume mounts are already present. Do not paste sample .env secrets or rerun secret generators to finish a fresh installation. A legacy variable description saying “CHANGE THESE” does not require replacing the generated keys. Keep the generated keys intact; rotating encryption keys against an existing database requires a recovery plan.

The template supplies the current worker command (pnpm run start in /app/apps/worker), matching the application release, and exposes the app on port 5560. The worker performs first-install database upgrades automatically; allow it to finish before using the app. No separate manual worker command or Docker Compose installation is needed. Database and auxiliary-service connections use private service references; use only the app domain for the browser. The saved Postgres and Redis services also have public TCP proxies: review whether you need that external access before production use.

Common Use Cases

Collect and inspect LLM traces and application metadata.
Explore quality, cost and latency using the configured application features.
Connect evaluation providers and process background tasks.

Dependencies for LangWatch (AI Observability) Hosting

The application and worker are aligned to LangWatch 3.22.0 and pinned by immutable image digest. NLP and evaluation helper images are also pinned to their tested digests. ClickHouse is configured as 25.10.2.65, Redis as 8.2.1, and PostgreSQL as the 18 image channel. Pinned application/helper digests keep fresh deployments on the validated image set. PostgreSQL remains on its existing maintained image channel. Updating this template does not update running customer projects.

Deployment Dependencies

A Railway account with resources for all seven services and their volumes.
Your own credentials for optional model/evaluation and email providers.
LangWatch source and license, documentation, and authentication configuration.

Why Deploy LangWatch (AI Observability) on Railway

The template groups the application, workers and supporting services in a single project with generated configuration. You retain responsibility for access, retention, backups, resources and version compatibility.

Operations, costs and validation scope

Review the saved DISABLE_PII_REDACTION=true setting before sending personal or sensitive data. Configure sampling and retention, and avoid putting credentials in traces. Keep database backups and generated encryption/session keys in a protected recovery plan; volumes are not backups.

This seven-service stack has baseline usage even while idle, with additional cost from traffic, evaluations, databases and retained telemetry. Provider charges are separate.

Release and fresh-install validation: 3 November 2026. Validation covers first signup, authenticated login, organization/project creation, a project-scoped trace-only API key, authenticated OTLP trace ingestion and stored trace readback, rejection without credentials, and trace persistence after a ClickHouse container replacement. Optional paid model-provider evaluations and outbound email were not tested. Existing projects require a separate upgrade and backup plan; they do not automatically receive template changes.