Developers
Build on Auvryn
Everything in Auvryn is available through a REST API described by a public OpenAPI 3.1 document, and through a TypeScript SDK, a CLI and an MCP server published on npm. Executions are asynchronous: you create an ExecutionJob, follow its status, then read its verified result, provenance and telemetry.
Auvryn is provider-neutral execution infrastructure for AI models. It validates ONNX models, runs them as durable ExecutionJobs on infrastructure it does not own, and returns verified results with provenance and telemetry.
The Developer API
- Base URL
- https://api.auvrynspace.com/api/v1
- Description
- OpenAPI 3.1, public: https://api.auvrynspace.com/api/v1/openapi.json
- Authentication
- API keys sent as a Bearer token. Keys carry scopes, can be limited to one project, expire, and are stored only as hashes.
- Creates
- Idempotent: send an Idempotency-Key header and a retried request never creates a second resource.
- Errors
- One error envelope with a stable code and a request ID; resources of another workspace answer “not found”.
- Rate limits
- Per API key: a 429 carries Retry-After. Workspaces also have limits, such as concurrent runs and daily sample runs.
The asynchronous flow
- Create a model and a version, then upload the ONNX file to the signed URL the API returns.
- Request validation; optionally run a sandbox test.
- Optionally define an area of interest and ask a provider for a cost estimate.
- Create an ExecutionJob for a provider; poll it until it is final.
- Read the result: its artifacts, their SHA-256, and the execution’s provenance and telemetry.
Interfaces
Four ways in, one API underneath: the same API key, scopes, workspace limits and provider visibility apply to all of them.
- REST API
- Public. https://api.auvrynspace.com/api/v1, described by OpenAPI 3.1: https://api.auvrynspace.com/api/v1/openapi.json
- TypeScript SDK
- npm install @auvryn/sdk (Node.js 24+, ES modules): https://www.npmjs.com/package/@auvryn/sdk
- CLI
- npm install --global @auvryn/cli: the auvryn command, JSON output and documented exit codes for CI. https://www.npmjs.com/package/@auvryn/cli
- MCP server
- npx -y @auvryn/mcp: a local stdio server; agents operate through the same authenticated and permissioned API, with writes and execution behind explicit settings. https://www.npmjs.com/package/@auvryn/mcp
- Source
- Apache-2.0, published from https://github.com/auvryn/developer-tools with npm provenance.
Quickstart
- Sign up and create a workspace and a project in the web app.
- Create an API key in Workspace → Developer → API keys; it is shown once.
- Set AUVRYN_API_KEY and AUVRYN_PROJECT_ID, then install the SDK or the CLI.
- Run the sample: auvryn executions sample --wait, or auvryn.executions.createSample({ projectId }) then auvryn.executions.wait(…).
- Read the result and the job’s provenance and telemetry: the sample is simulated and says so (simulated: true).
- Bring your own ONNX model: upload, validate and test it in the sandbox through the same interfaces.
Execution environments
Interfaces are public; execution environments are listed per workspace. GET /api/v1/providers returns only the providers your workspace can use.
- Sample Environment
- Public, for every workspace. Simulated: no model runs on real hardware, and provenance says simulated: true. A few runs per workspace per day.
- Real cloud execution
- Real and verified in production, for an internal workspace only; not offered to customers yet.
- Real edge execution
- Real, under internal evaluation; not offered to customers yet.
- Orbital execution
- Not yet: no orbital execution has happened through Auvryn.
For agents and tools
Auvryn publishes what a program needs to discover it without reading prose.
- /.well-known/api-catalog
- The API catalog (RFC 9727): where the API, its OpenAPI description and its documentation live. https://auvrynspace.com/.well-known/api-catalog
- /capabilities.json
- A small, versioned description of what Auvryn supports today: model formats, interfaces, execution model, environments and whether each is simulated or real, provenance and telemetry, availability and restrictions. https://auvrynspace.com/capabilities.json
- OpenAPI 3.1
- https://api.auvrynspace.com/api/v1/openapi.json
ONNX model guides
- How to validate an ONNX model before deploymentChecker, opsets, inputs and outputs, a test load in ONNX Runtime and a parity check.
- Why an ONNX model gives different resultsPyTorch vs ONNX Runtime, CPU vs GPU: what is expected and how to find the cause.
Current limits
- Model format: ONNX.
- Execution: asynchronous jobs only; there is no synchronous inference endpoint.
- Providers: you choose the provider for each job; Auvryn does not pick one for you.
- Access: early access; there is no public pricing yet.
Related
Facts reviewed 2026-10-08.