SDK capabilities and limits
Understand what is a context read, a prepared model and an executable operation.
Available surfaces
| Surface | Use | Important limit |
|---|---|---|
| Workspace context | Discover, read business understanding and outcome guidance | Bounded pages; context-read consent and current workspace access required |
| Published models and processes | Read versions, stage rules and supported process run state | Publication and separate grant required; no universal ERP execution |
| Model setup and import | Recommend setup; prepare and commit supported definition imports | Apply/commit requires context-write and current manager rights; prepared is not saved |
| Computations | Submit supported private calculations and read durable job results | Separate computation grant; queued is not complete; new results need application-level assembly |
| Existing capability integrations | Invoke declared operations and retrieve authorized files | Only where installed, qualified and separately consented |
Keep access narrow
The read-only OAuth example requests toolkits:read and workspace_context:read. The legacy toolkits:read name remains a transport prerequisite; it does not itself authorize execution or every published capability.
Model changes require workspace_context:write. Computations use computations:read and, for running, computations:run with a separate workspace computation grant. Existing process and file access retain their own grants. Never request all scopes just because the server lists them.
Provider use
Ordinary discovery and definition reads do not invoke an LLM. Explicit model setup, model authoring, hosted guidance or computation planning may use Neuron’s configured provider. Keeping the conversation in your application does not mean every supported tool is provider-free.
What you must still handle
New model-setup, import, provider-free process-operation and computation operations have generated request types and schemas through client.call. Their responses are structured JSON; they do not all have high-level SDK receipt validation or complete result assembly. Follow the server contract, completeness markers and result status before showing a conclusion.
The SDK is not an agent runtime, an automatic database mapper, a scheduler for arbitrary jobs or a substitute for application authorization. Source credentials stay inside Neuron.