Smarter AI Resource Reference

Smarter is a governed platform for building LLM-powered applications. Every capability in Smarter — from calling a model, to running a tool, to storing a credential — is expressed as a declarative Resource.

What sets Smarter apart is how those Resources are created: entirely through a Smarter Application Manifest (SAM), a plain YAML file that declares a Resource’s desired state. There is no separate SDK to learn and no resource-specific API call to make — an Account, a Secret, an LLMClient, an Orchestrator, a Plugin, all of it is created and managed the same way, by applying a manifest. No other platform offers this: the entire surface area of the system, from infrastructure concerns like credentials and self-hosted models down to individual conversational guardrails, is addressable through one declarative interface. Because SAM files are versioned, diffed, and reviewed like any other infrastructure-as-code, and because they’re plain, human-readable YAML, they can be understood and applied by non-developers using the web console.

Example Manifest
apiVersion: smarter.sh/v1
kind: Secret
metadata:
  description: |
    An example secret manifest. This is a real credential used for authenticating to
    the Smarter Remote SQL test database. This is a public-access test database
    used for development and testing purposes only.  Do not use this credential
    for production purposes. Password value is the same as the user.
    See the annotations below for connection details.
  name: smarter_test_user
  version: 1.1.0
  tags:
    - example
    - test
    - stackademy
  annotations:
    - smarter.sh/test-db/host: sql.lawrencemcdaniel.com
    - smarter.sh/test-db/port: "3306"
    - smarter.sh/test-db/db: smarter_test_db
    - smarter.sh/test-db/user: smarter_test_user
    - smarter.sh/test-db/pwd: smarter_test_user
    - smarter.sh/test-db/access: anyone can access this test database
spec:
  config:
    description: smarter_test_user password. Password value is the same as the user.
    value: smarter_test_user

Smarter manifest files can be processed by the Smarter CLI as well as by drag-and-drop on the web console, both of which interact seamlessly and securely with the Smarter API to provision, update, and manage the lifecycle of Smarter Resources. The Smarter Project maintains a VS Code Extension that provides syntax highlighting, manifest validation, and inline documentation for authoring SAM files.

Resources compose into four layers, and understanding those layers is the fastest way to understand the platform:

Access & Governance

Smarter Account and Smarter Secret establish who is acting, what they’re permitted to access and spend, and how credentials are stored and retrieved without ever appearing in plaintext. Every other resource operates within the boundaries these two sets.

Model Connectivity

Smarter Provider and Smarter LLM Host are the two ways a model reaches Smarter — a third-party API such as OpenAI or Anthropic, or a self-hosted, freely downloadable model you run and manage yourself. Smarter LLM Client sits in front of both, giving prompts and applications one consistent interface regardless of which kind of model is actually answering the request. Smarter Orchestrator sits in front of LLMClient in turn, coordinating a collection of LLMClients — as Harnesses, each with its own role — toward a single agentic objective, using a configurable strategy such as sequential hand-off, parallel fan-out, supervisor/worker delegation, routing, or voting/debate.

Extensibility & Trust

Smarter Plugin is how a model does things beyond generating text — Static, Skill, API, and SQL tool calls — reaching external systems through Smarter Connection and the wider Model Context Protocol ecosystem through Smarter MCP Client. Smarter VectorStore gives those plugins and prompts a self-hosted knowledge base to retrieve from. None of this is unsupervised: Smarter Guardrail inspects every request and response, enforcing moderation and security policy regardless of which model, plugin, or data source is involved.

Conversation

Smarter Prompt is where it all lands — a chat conversation with archivable history, capturing what was asked, what tools ran, what the guardrails did, and what came back.

This layered design is what makes Smarter auditable end to end: an administrator can trace any single Smarter Prompt back through the guardrail checks, tool calls, and model connection that produced it, and back further still to the account and budget that authorized it in the first place. Where a prompt was produced as one step of a multi-agent workflow, that trail extends one layer further still, to the Smarter Orchestrator run and the sequence of Harness invocations that led to it.

Because Resources are layered rather than monolithic, different roles on a team naturally gravitate toward different parts of the stack, without needing to understand the whole thing to be productive:

Technical Reference