Dvara Examples

Reference configurations, compose files, and SDK integration samples for the Dvara LLM Gateway.

Latest release: 1.8.1 — compatible with Dvara 1.8.1.

Contents

Directory	Description
docker-compose/	Ready-to-run Docker Compose stacks (quick-start, multi-provider, full, ollama, with-email)
getting-started/	First-request scripts in Python and Node.js — basic chat, streaming, structured outputs, multi-provider
sdk-integrations/	Framework examples — OpenAI SDK, LangChain, LiteLLM, Pydantic AI, Vercel AI, Spring AI, LangChain4j

Quick start

git clone https://github.com/dvarahq/dvara-examples.git
cd dvara-examples/docker-compose/quick-start
cp .env.example .env
# edit .env — set OPENAI_API_KEY. Leaving DVARA_LICENSE_KEY blank still runs
# everything except the MCP and A2A planes; a DVARA- envelope activates those two.
docker compose up -d

Gateway ready at http://localhost:8080, DVARA Console at http://localhost:8090.

# Create a tenant and API key in the Console, then:
export DVARA_API_KEY="gw_<your-key>"

cd ../../getting-started/python
pip install -r requirements.txt
python dvara_test.py

Compatibility

Each release of these examples is pinned to a specific Dvara LLM Gateway version. Use the matching examples release for your gateway version.

Examples release	Compatible Dvara version
1.8.1	Dvara 1.8.1
1.8.0	Dvara 1.8.0
1.7.0	Dvara 1.7.0
1.6.0	Dvara 1.6.0
1.5.0	Dvara 1.5.0

Older examples releases are listed on the releases page.

Changelog

1.8.1

• Version bump — every Compose stack, Kubernetes/Helm recipe, the DigitalOcean recipe, and the jbang GKE tooling now pin 1.8.1 (chart oci://ghcr.io/dvarahq/charts/dvara:1.8.1).
• Drop-in over 1.8.0. Nothing in these recipes changes shape: no renamed image, no new service, no moved port.
• One setting can stop the gateway starting. If you turned the ML classifier on and set provider=onnx-injection, 1.8.1 removed that provider and the gateway refuses to start. Use provider=defender, whose model is in the image and needs no volume. None of these stacks sets it.

1.8.0

• Version bump — every Compose stack, Kubernetes/Helm recipe, the DigitalOcean recipe, and the jbang GKE tooling now pin 1.8.0 (chart oci://ghcr.io/dvarahq/charts/dvara:1.8.0).
• Every /v1 call needs an API key. Dvara 1.8.0 removed DVARA_LLM_GATEWAY_REQUIRE_API_KEY, so the Compose stacks no longer set it. Setting it to false on 1.8.0 stops the gateway starting.
• Set DVARA_AUDIT_HMAC_SECRET before running a production profile. Dvara 1.8.0 no longer generates the audit signing secret. These recipes run the default profile, where the gateway starts without one; a production profile refuses to start until every gateway and Flightdeck process has the same value.

1.7.0

• Version bump — every Compose stack, Kubernetes/Helm recipe, the DigitalOcean recipe, and the jbang GKE tooling now pin 1.7.0.
• dvara-llm-gateway is now dvara-gateway. The image was renamed; the old package still resolves to pre-release images, so repoint rather than leaving it.
• full/ lost two containers, and they are not coming back. dvara-mcp-gateway (8070) and dvara-a2a-gateway (8075) are retired images — the MCP and A2A planes now run inside the gateway process. The paths are unchanged and moved to the gateway on 8080. Repoint anything addressing :8070 or :8075.
• The Community / Enterprise image split is gone. There are no -ee variants and no private packages: there is one artifact per application, all public, and a licence decides what runs. Unlicensed installs get policies, PII, guardrails, audit, budgets and cost attribution; a DVARA- envelope additionally activates the MCP and A2A paths. Nothing refuses to boot without one.
• full/ is folded into quick-start/. Once the planes moved into the gateway it was the same three services plus a DVARA_LICENSE_KEY, so the directory named a topology that no longer existed. Run quick-start/ and set a DVARA- envelope to get what full/ gave you.

1.6.0

• Version bump — every Compose stack, Kubernetes/Helm recipe, the DigitalOcean recipe, and the jbang GKE tooling now pin 1.6.0 (chart oci://ghcr.io/dvarahq/charts/dvara:1.6.0).
• Documented the Community / Enterprise image split. Since 1.5.0 the images are in two registries' worth of visibility: dvara-llm-gateway and dvara-flightdeck are public, while dvara-llm-gateway-ee, dvara-flightdeck-ee, dvara-mcp-gateway, and dvara-a2a-gateway are private to customers. Nothing in this repo said so, so the full/ stack — which uses all four private images — failed on docker compose pull with an opaque denial for anyone without registry access. full/ now has its own README stating the requirement up front, and the image table in docker-compose/README.md is split public / private.
• Corrected the full/ stack description — it has run five services since 1.3.0 (it gained the A2A plane then), but the variants table still listed four and omitted dvara-a2a-gateway.

1.5.0

• Version bump — all Docker Compose stacks (quick-start, multi-provider, ollama, full, with-email) now pin ghcr.io/dvarahq/dvara-*:1.5.0.
• License key now optional — .env.example ships with DVARA_LICENSE_KEY= blank: blank runs Community Edition; set a DVARA- signed envelope for Enterprise.
• Kubernetes image path fix — corrected the image repository to ghcr.io/dvarahq/dvara-llm-gateway / dvara-flightdeck (removed the stray /dvara/ path segment) across the DOKS, GKE, multi-region, and single-tenant values.

Older changes are in the notes of each release on the releases page.

Documentation

Full product docs at dvarahq.com/docs.

License

MIT — see LICENSE.