rulvarAn embeddable TypeScript engine for multi-agent LLM workflows.

Durable by construction: a completed LLM call is never paid for twice. Three orchestration modes on one runtime, one journal, one budget path. Hard dollar ceilings, vendor-neutral provider adapters, VCR cassettes for hermetic tests, and ModelKnowledge that remembers which models work in your project. No server, no database, no control plane.

Get started

What is rulvar?

View on GitHub

Never pay twice

The journal is a content-addressed memoizing log of completed effects. On resume, finished LLM calls replay from disk; inserting a new call costs exactly one live call.

The journal

Budgets as hard invariants

A per-run USD ceiling enforced at admission, before every agent turn, and by cutting live streams on crossing. Overshoot is bounded and documented; exhaustion returns partial results, never null.

Budgets and termination

Three orchestration modes

Human-written scripts, machine-written scripts with lint and self-repair, and a dynamic orchestrator agent - all on the same runtime, journal, and budget path.

Orchestration modes

Vendor neutrality by construction

The core imports no provider SDKs. Anthropic and OpenAI adapters ship first-class, an openaiCompatible factory covers compatible servers, and a bridge wraps any Vercel AI SDK model.

Providers

Multi-model at every level

Models resolve per invocation, not per agent - call override, agent profile, workflow defaults, engine defaults. Invocation roles route planning, extraction, and summarization to different models.

Model routing

Testable by construction

FakeAdapter, VCR cassettes with secret redaction, replay-strict runs, and vitest matchers make agent workflows testable without a single live call in CI.

Testing

Observability out of the box

One typed event stream for every model call, tool call, spawn, and budget decision, with OTel export, cost reports, and a queryable run handle.

Observability

Embeddable first

A library, not a platform - it lives inside your Node.js app. The CLI, HTTP server, and queue worker shells are optional and build strictly on the public API.

Shells

ModelKnowledge

A per-project knowledge base of model behavior - verified facts, eval evidence, decay, and a rendered knowledge card that teaches orchestrators which models to spawn.

ModelKnowledge

v1.1.0Apache-2.0