Build on LibraOS
Everything the product does is reachable over HTTP, and everything in this guide is something an application can do without a change to the server. This page is the entry point for building one.
The SDK
pip install libraos-sdk
The Python SDK is Anthropic Managed Agents compatible, so code written against that shape works here, with LibraOS's multi-model routing and employee features available on top. Point it at a deployment:
export LIBRA_OS_URL=https://your-deployment
export LIBRA_OS_API_KEY=...
Source, examples and the partner integration docs: github.com/libraos/sdk
The API contract
The SDK is generated from a published OpenAPI specification, and so is everything else — the Go client, the TypeScript types, the CLI. If you would rather call the API directly, or generate a client for a language we do not ship, the spec is the authority:
openapi/libra-os-partner.v1.yaml
It is worth reading even if you use the SDK, because it is where response
fields are defined — including the ones that tell you whether an answer can be
trusted. grounding, retrieved_chunks and citations_verified are described
in Workspaces & memory and Web search; the
spec is where their exact values live.
What is available today
| status | |
|---|---|
Python — pip install libraos-sdk | published |
| OpenAPI spec | published, the source of truth |
| CLI | in the SDK repo |
| TypeScript client | in the repo, not published to npm — clone and build from clients/typescript |
| Go client | in the repo under cli/internal/client |
The TypeScript client is generated and tested but has no npm release yet. Use
it from a checkout rather than expecting npm install to work.
Worked examples
The Cookbook learning paths are the main entry point for runnable Runtime tutorials: First Agent, independent Knowledge/Tools/Memory branches, Async Recovery, and complete applications. Each path identifies prerequisites, observable outcomes and failure checks. Read the recipe README alongside its source; some examples need companion services or multiple user identities.
The SDK repository also carries examples:
examples/simulator— a synthetic-customer simulator for exercising an agent against generated conversationsexamples/react-ui-demo— a front end talking to a deploymentemployees/— employee definitions you can install and adapt, including theemail-classifierused in Ticket routing
Two end-to-end guides walk the whole path: Customer support and Ticket routing.
What the platform handles for you
Worth knowing before you build around it, because these are not things an application needs to implement:
- Grounding and citation checking — every answer carries a verdict; you read it rather than compute it
- The AI firewall — inbound and outbound screening on every turn
- Agent orchestration — planning, parallel skill execution, circuit breakers and per-stage deadlines
- Knowledge retrieval — collections, bindings and scope enforcement
- Approval workflow — the dry-run and pending-action gate for side-effecting tools
An application decides what to ask for and what to do with the answer. Retrieval, safety and orchestration are the platform's job.
Before you go to production
- System architecture — the boundary between your application, the runtime, its stores, model services, and external tools.
- Portability contract — what behaves identically across deployment shapes and what degrades. Read this before assuming a capability is present.
- Security model — what leaves the deployment.
- Managing memory — what persists, and how to scope it to the right person.
Where to go next
- Getting started — a running deployment in about six minutes
- Creating an agent — your first agent
- Employee YAML — the definition format