> ## Documentation Index
> Fetch the complete documentation index at: https://docs.infercrane.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Integration ownership and licenses

> Understand what InferCrane owns, what specialist systems retain, and how new adapters qualify.

# Compose specialist systems without blurring ownership

InferCrane integrates with established infrastructure and inference projects through narrow,
replaceable contracts. A compatible license is necessary but not sufficient. Every integration also
needs a pinned version, a security review, conformance tests, explicit mutation ownership, and
real-system qualification before it can be advertised as supported.

| System category  | Integration model                                              | Candidate upstreams                                                                                                                                    | License boundary                                          | InferCrane owns                                     | Specialist system retains                               |
| ---------------- | -------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------- | --------------------------------------------------- | ------------------------------------------------------- |
| LLM gateway      | OpenAI-compatible target or replaceable gateway adapter        | [LiteLLM](https://github.com/BerriAI/litellm), [Envoy AI Gateway](https://github.com/envoyproxy/ai-gateway)                                            | Review exact distributed paths; no fork is bundled        | Endpoint identity, budgets, evidence, releases      | Provider translation catalog and upstream credentials   |
| Training         | Signed checkpoint handoff; job adapter only after demand       | [Kubeflow Trainer](https://github.com/kubeflow/trainer), [MLflow](https://github.com/mlflow/mlflow), SkyPilot jobs                                     | External service and artifact terms remain operator-owned | Artifact lineage, qualification, deployment handoff | Training data, execution, storage, and scheduling       |
| Vector database  | Dependency identity and health evidence only                   | [Qdrant](https://github.com/qdrant/qdrant), [Milvus](https://github.com/milvus-io/milvus)                                                              | External service; not redistributed                       | Endpoint linkage and evidence references            | Vector storage, indexing, and retrieval                 |
| Agent framework  | Stable endpoint, SDK examples, telemetry, and request identity | [LangGraph](https://github.com/langchain-ai/langgraph), [Google ADK](https://github.com/google/adk-python)                                             | External application dependency                           | Inference identity and operational evidence         | Planning loop, tools, and application state             |
| Sandbox          | External identity plus scoped, expiring inference access       | [E2B](https://github.com/e2b-dev/E2B), [Kubernetes Agent Sandbox](https://github.com/kubernetes-sigs/agent-sandbox)                                    | External runtime and credentials                          | Endpoint access policy and evidence references      | Isolation, commands, files, snapshots, and networking   |
| Inference engine | Runtime Contract adapter and conformance                       | [Ollama](https://github.com/ollama/ollama), [llama.cpp](https://github.com/ggml-org/llama.cpp), [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM) | Audit the exact engine and redistribution path            | Capability truth, lifecycle intent, and evidence    | Engine internals, kernels, decoding, and distributed KV |
| Kubernetes       | Standard APIs against an operator-owned cluster                | Kubernetes, KServe, Kind for tests                                                                                                                     | InferCrane does not distribute a cluster                  | Workload intent, manifests, lifecycle, and evidence | Cluster, networking, storage, devices, and scheduling   |
| GPU scheduler    | CRD or API selected by the cluster owner                       | [Kueue](https://github.com/kubernetes-sigs/kueue), [Volcano](https://github.com/volcano-sh/volcano)                                                    | External cluster component                                | Serving intent and scheduling evidence              | Queueing, placement, and device allocation              |
| Workflow engine  | Events, webhooks, SDK steps, and bounded async inference       | [Argo Workflows](https://github.com/argoproj/argo-workflows), [Temporal](https://github.com/temporalio/temporal)                                       | External application service                              | Inference operations and inference idempotency      | Generic DAG and workflow execution                      |

## Build, integrate, or reject

Build a capability only when it belongs to InferCrane's durable inference contract and no specialist
system already owns it. Integrate when a mature upstream project owns the concern. A fork is
acceptable only when all of these conditions are true:

1. The license permits the exact distribution model.
2. An adapter cannot produce the required user outcome.
3. The divergence is deliberately small and bounded.
4. A named maintainer owns upstream security merges.
5. A removal or upstreaming plan exists.

No candidate listed above currently meets that fork threshold.

## Qualification ladder

Each adapter advances independently:

```text theme={"theme":"css-variables"}
REGISTERED → CONTRACT_TESTED → LOCAL_QUALIFIED → REAL_SYSTEM_QUALIFIED
```

Registration means a capability is discoverable. It does not prove credentials, networking,
capacity, runtime behavior, performance, billing, cleanup, or recovery on a real system. The CLI,
API, console, and documentation must display the achieved tier rather than infer support from the
presence of an adapter.