Clients

Vectros is a REST platform, but you almost never call it with raw HTTP. Four client surfaces meet you where you are — from a typed SDK in your application code, to a one-command provisioning CLI, to an agent-native MCP server, to a declarative blueprint format that stands up a whole app's data model from a file. Pick the one that fits the job; they all talk to the same API and inherit the same tenant isolation and scope rules.

SurfaceWhat it isReach for it when…Doc
SDKTyped client libraries generated from the API spec. Node is the verified reference; Java and Python are also generated.You are writing application code and want types, pagination helpers, and streaming handled for you.sdk.md
CLI (@vectros-ai/cli, vectros)A command-line tool for provisioning: blueprints, app contexts, identities, scoped keys, access bindings, and a one-shot MCP credential bootstrap.You want to scaffold or provision from a terminal or CI — without writing code or handling your root key directly.cli.md
MCP server (@vectros-ai/mcp-server)A Model Context Protocol server that exposes the Vectros data plane to AI agents as 19 callable tools.You want an agent (Claude Desktop, Cursor, and others) to read, write, search, and reason over your tenant's data with no integration code.mcp.md
Blueprints (@vectros-ai/blueprints)A declarative app-model-as-config format: schemas, an access profile, a service principal, seed data, and roles, all in one file.You want to describe an app's data model once and provision it idempotently.blueprints.md

How they fit together

The pieces compose into a no-code path from a file to a working, agent-driven app:

  1. A blueprint declares the schemas, the least-privilege access profile, a service principal, optional seed records, and optional roles.
  2. The CLI bootstrap command reads the blueprint, provisions everything it declares, and mints a narrow scoped key (ssk_*) — merging it straight into your MCP client config.
  3. The MCP server picks up that key and gives an agent the data-plane tools.
  4. Your application code — or a forked reference app — uses the SDK for the same operations programmatically.

A blueprint can only ever request data-plane access (records, schemas, search, documents, folders, inference). The CLI is the trust boundary: it hard-refuses any control-plane scope, so even an untrusted, hand-written, or agent-authored blueprint cannot escalate the credential it provisions.

A note on versions

The API is specified once and the Node SDK is generated from that spec; the spec is currently at 0.27.0. The clients in this repository do not all track the very latest spec in lockstep — a few capabilities (record/document/folder PATCH, and the typeName-or-schemaId choice on record creation) need a 0.26 client. The CLI and MCP server already bundle a 0.26 staging build, so they have them; the React toolkit and the reference web apps still pin a 0.23 staging build, so those capabilities are not reachable through the reference apps' UI. Each doc calls out exactly which surface needs the newer SDK; see sdk.md for the full picture.

Where to go next

  • sdk.md — install, construct a client, and make your first call.
  • cli.md — the full command catalog and the blueprint lifecycle.
  • mcp.md — the 19 tools, both transports, and the credential bootstrap.
  • blueprints.md — the blueprint format reference.
  • The blueprint walkthroughs (getting-started, clinical-intake, coding-agent-memory, second-brain) — end-to-end, narrated builds on top of these clients.