Why Arivie

Arivie exists because teams want agentic analytics inside their own application without rebuilding Mastra primitives, duplicating MCP handlers, or inventing a parallel agent loop. The framework name, license, and foundation are locked in the project pivot document: Arivie ships under Apache-2.0 for v0.x, published as @arivie/*, and is built on Mastra — not beside it.

Built on Mastra, not around it

Every Arivie package whose responsibility overlaps a Mastra primitive extends that primitive’s interface (REQ-1). Storage uses @mastra/pg; the filesystem-driven explore tool uses Mastra Workspace; embeddings use Mastra Vector; the agent uses Mastra Agent; MCP uses @mastra/mcp. Arivie does not run a parallel agent loop (REQ-2). AI SDK remains the escape hatch for narrow operations Mastra has not surfaced yet, but the production architecture treats Mastra as the loop owner.

This matters for maintainers: when Mastra ships observability, dev TUI, eval framework, or MCP server primitives, Arivie consumes them rather than forking. Where Mastra does not cover data-agent-specific concerns — semantic-layer linting, cross-mode coverage, registry UI — Arivie ships those pieces in @arivie/semantic, @arivie/registry, and related packages.

Single-tenant by design

The pivot’s central simplification is one Arivie instance = one owner (one company’s data, one tenant’s data, or one department’s data). The instance scope is the boundary. User-level permissions inside an instance are pass-through to the owner’s database role: if the configured DB role cannot SELECT a table, the agent cannot either. Multi-tenant SaaS embedders deploy N instances, one per tenant, using @arivie/deploy recipes rather than in-process tenant routing.

What was removed from the architecture is explicit in the pivot: tenant-scoped views with session variables as a firewall, in-process multi-tenant mode, cross-tenant eval probes, and a resolveTenant(req) contract. Those are replaced by owner-scoped instance config plus resolveUser(req) for personal memory and pass-through ACLs.

v0.1 is production-shaped

The pivot commits to shipping the production deliverable list at v0.1 — documentation, eval suite, MCP parity, all three context modes, deployment recipes, and registry UI — with no deferring core primitives to a later milestone. RFC-002 encodes the same bar through REQ-30–32 (dogfood eval ≥ 90%, BIRD-Interact score, bundle and latency budgets).

When to choose Arivie

Choose Arivie when you already run Postgres analytics data, want agents that read your warehouse through a curated semantic layer, need HTTP and MCP on the same Mastra Agent, and can map each customer or org to its own deployed instance. If you need arbitrary multi-tenant routing inside one process, Arivie is the wrong fit — deploy multiple instances instead.