Deploy Agenta LLM Engineering

Deploy and Host Agenta LLM Engineering on Railway

LLM evaluation and tracing with databases, workers and object storage.

Validation scope: static configuration checks only. Image builds and Railway application workflows have not been validated. Deploying this template incurs Railway usage and any external provider charges.

About Hosting Agenta LLM Engineering

Thirteen services adapted from the upstream Railway OSS layout: PostgreSQL, Redis, SeaweedFS 4.37, web/mobile, API/services/runner, two workers, cron, Supertokens and gateway. API migrations gate background startup, and its signing key persists on /data. SeaweedFS is pinned to the upstream-required STS-compatible version. Runner execution uses external Daytona; host Docker/FUSE is unavailable. No model or sandbox credits, enterprise features, HA or fixed hosting cost are included. The OSS code is MIT-licensed; enterprise code has separate terms. This is the largest and highest-validation-effort draft in this batch.

Setup

Enter runner.AGENTA_RUNNER_DAYTONA_API_KEY for your own Daytona account. Open agenta using gateway username admin and agenta.ACCESS_PASSWORD, then create the application account. Configure a model provider in Agenta. External SDK calls require X-Template-Key with the gateway password in addition to the native API key; confirm your client supports this.

Deployment Dependencies

Operator inputs

Generated credentials: postgres.POSTGRES_PASSWORD, redis.REDIS_PASSWORD, api.AGENTA_AUTH_KEY, api.AGENTA_CRYPT_KEY, api.AGENTA_SERVICES_INTERNAL_KEY, api.AGENTA_RUNNER_TOKEN, api.AGENTA_STORE_ACCESS_KEY, api.AGENTA_STORE_SECRET_KEY, api.AGENTA_STORE_SIGNING_KEY, agenta.ACCESS_PASSWORD. Never copy another deployment's secrets. Preserve encryption/signing keys with backups.

Why Deploy Agenta LLM Engineering on Railway

The definition connects private dependencies, declares persistent volumes and exposes the listed application endpoints through Railway HTTPS. Repository adapters build from codex/fifteen-template-drafts with the repository root as Docker context. Configure one replica for each stateful service.

Common Use Cases

LLM evaluation and tracing with databases, workers and object storage. Use a separate test project to verify the workflow below before putting real data into the instance.

Dependencies for Agenta LLM Engineering

Reviewed upstream release: v0.118.0. Image digests are recorded in images.round6.lock.json; source receipts are in sources.round6.lock.json. A matching version label does not independently prove image-to-source provenance. Base-image pins do not lock later apt/apk or language-package resolution.

Recommended acceptance checks

Build adapters; initialize all three databases; run core/tracing migrations; verify worker readiness; sign up and sign in through Supertokens; evaluate a prompt; ingest and view a trace; save and retrieve a stored artifact via STS; run a small Daytona task; verify native API auth plus owner header; restart all roles; restore databases, SeaweedFS and signing/encryption keys together.

Operations and recovery

Back up every listed persistent volume and export every application database, preserving the generated encryption/signing secrets. Restore into a separate test instance and repeat the functional workflow before relying on the backup. Keep internal databases private. An owner gateway /healthz response proves only that gateway is alive, not that its application or dependencies are ready. After changing a public domain, update all application origin and callback settings. Review upstream migrations before upgrading; record the old source/image pins and a recoverable backup. No scheduled backup service is configured by this draft.

Validation status

The checked-in catalog supports static pin, reference, Docker COPY-path and script-syntax validation. Container builds, native proxy behavior, fresh Railway deployment, provider integrations, persistence and restore remain release gates. See RELEASES-ROUND-6.md for the publication record.