Gateway and release guide

Start with a published release

This guide was checked against Base v0.3.18 on 30 September 2026. The public repository provides versioned charts, release compositions, defaults, and validation tooling. It does not provide a complete cluster configuration or client secrets.

1. Review the release README, package selection and prerequisites, and release manifest.
2. Prepare the client-owned cluster configuration and credentials. The release lists Flux, required controllers/CRDs, ClickHouse operator, OpenBao setup tooling, and a DocumentDB CPU baseline among its prerequisites.
3. Follow the release’s adoption and migration requirements. This version requires separate API-key bridge and Studio credential-sync stages, with readiness dependencies. It also replaces Langfuse storage without migrating previous trace data.
4. Confirm selected packages, supported policy versions, endpoint access, and model pricing before connecting clients. Consult the published releases when selecting a newer version.

There is no verified one-command installer in this public release. Follow the versioned requirements with your deployment operator; a chart or source directory alone is not an installation or availability guarantee.

Product scope and published packages

The website describes Base as the AI gateway and its control, visibility, and runtime foundations. The published package catalog is not yet fully aligned with that product boundary.

• Integrated applications: the default application composition includes LibreChat core and Dify. Their databases, caches, search services, storage, and workers support those applications. Dify’s upstream workflow and agent services remain Dify features.
• Deployment-selected services: Docling and WireGuard are opt-in packages in v0.3.18. Docling supports gateway attachment inspection; the released WireGuard configuration provides a scoped Forgejo HTTPS gateway.
• Known mismatch: v0.3.18 still lists Forgejo as an included optional package, enabled only through client selection. Content Git hosting belongs to the separate knowledge-product positioning; its presence in this release does not establish a completed package migration.
• Source is not a shipped feature: LibreChat RAG and Code Interpreter directories exist, but the v0.3.18 release configuration explicitly excludes them. They are not advertised here as included Base capabilities.
• Client deployment needs confirmation: the public release does not reveal which optional packages a particular installer or client configuration selects. Confirm the effective selection with the deployment owner.

Private agent and knowledge add-ons have their own product, data, release, and support boundaries. This website update does not change released packages or announce availability of those add-ons.

Connect your tools

Use whatever frontend fits your team, provided its requests match the gateway’s supported API and enabled policies. LibreChat and Dify are integrated applications; other clients connect through standard APIs.

1. Obtain your deployment’s gateway API address, a validated Base API key, and a model ID approved for that caller from your administrator.
2. Set the client’s API endpoint/base URL and API-key fields. The reviewed release serves supported LLM requests through /v1/chat/completions and model discovery through /v1/models. Follow the client’s URL convention to avoid duplicating /v1.
3. Select an approved model and verify the actual request path, required features, streaming behavior, and policy results with your deployment.
4. For MCP clients, use the configured MCP endpoint with a transport and authentication method supported by the deployed Base version. LLM API settings do not establish MCP compatibility.

Example: n8n

For compatible LLM calls, configure n8n’s API endpoint/base URL and API-key settings to use Base, then select an approved model. Use a node/version that supports a custom endpoint and Chat Completions. See the upstream OpenAI Chat Model and API-key credential documentation.

This is a standard API connection, not a Neurwerk connector or certification. It does not establish compatibility with every n8n feature, the Responses API, or provider-hosted tools.

An API key identifies its validated caller. A shared service key does not automatically identify each end user or add SSO. Base does not replace application-specific permissions. Only supported requests routed through the gateway receive its controls.

Policy and visibility

Base includes Studio and monitoring. Studio provides supported administration, policy tools, AgentGateway request-log usage views, and logs. Prometheus/Grafana and Alertmanager cover metrics and alerts; Fluent Bit/OpenSearch cover logs; Langfuse supports LLM/MCP trace investigation.

AgentGateway computes measured model cost from recorded usage and configured pricing. The reviewed Base defaults supply no pricing catalog. Missing pricing must not be treated as zero spend. Hard budget enforcement is future work dependent on AgentGateway support.

PII protection depends on supported formats and configured policies, including any enabled attachment inspection. It does not guarantee perfect detection or regulatory compliance. External model destinations can receive data outside your infrastructure.

Source and licensing

The public Base repository licenses its owned content under MIT. Its README identifies that content as the versioned platform contract, not application source code. Third-party components and integrated applications retain their upstream licenses; consult the third-party notices and each component’s terms.

The repository license does not establish source availability or licensing for every referenced runtime image. Confirm those per artifact. Open-source claims on this site refer to Base, not the separate private add-ons.