Skip to content

GraphReFly

Reactive Graph Protocol · Causal Tracing

Graph+Re[active]=Fly

GraphReFly makes the future of human-AI collaboration real. One reactive graph where humans and AI work as peers — composing, reviewing, and modifying together. Changes propagate automatically; the model reruns only when it must. A thin harness layer lets you turn each black-box step into a verified outcome. Every flow observable, every change auditable. Step by step, probability becomes predictability.

Get started
GRAPHREFLY · REACTIVE GRAPH PROTOCOL · CAUSAL TRACING
stage 1 · connection lifecycle stage 2 · diamond resolution stage 3 · node anatomy stage 4 · four sugars stage 5 · graph level stage 6 · orchestrate · messaging stage 7 · agent loop · memory
START → DIRTY
→ DATA → COMPLETE
DIRTY converges
DATA fires once
cache · version
guard · up / down
state · producer
derived · effect
describe · observe
batch · subgraphs
pipeline + hub
static · dynamic
loop · memory
open substrate
graph.describe() graph.observe() batch()
Specifications

The numbers behind the protocol.

Runtime deps
0
Pure-TS substrate. No transitive surprises in your bundle.
universal · browser + node
Primitive
1
node with 4 sugars: state · producer · derived · effect.
one cell, every workflow
Operators
70+
transform · combine · buffer · window · rate-limit · retry · break
composable, type-safe, tree-shaken
Connectors
40+
from/to: Kafka · OTel · MCP · HTTP · Postgres · Redis · WebSocket · …
production-grade integrations
The Problem

Three harness gaps.

Every team building with LLMs hits the same three walls. They are not modeling problems — they are harness layer problems. GraphReFly solves all three in the same protocol.

01

Context without control.

Agent state is scattered across files, databases, and conversation transcripts. No single system sees it all. The agent becomes the integration layer — and fails.

GraphReFly

Reactive state harness.

A single reactive topology connects every source. State pushes to dependents automatically — no re-reading, no stale snapshots. One graph sees what ten dashboards can't.

Demo: PagerDuty Triage →
02

Action without explanation.

Your agent flagged something. Why? You check logs — a wall of text. You check state — a snapshot, not a story. You can't explain the decision, so you can't trust it.

GraphReFly

Causal tracing.

Ask "why was this flagged?" and get a structural causal chain — from source to conclusion. Not a log dump, but traceable causality that persists across sessions.

Demo: Spending Alerts →
03

Composition without guardrails.

LLMs compose agent workflows by writing arbitrary code. Error space is unbounded. One hallucination cascades through the pipeline. No policy enforcement, no safe rollback.

GraphReFly

Declarative GraphSpec.

Describe workflows in a declarative schema — like SQL for data flows. LLMs compose structurally, not imperatively. Policy enforcement, persistent checkpoints, atomic mutations built in.

Demo: Knowledge Graph →
The Harness

Four requirements.

A harness layer between your LLM and production must satisfy all four. Miss one, the whole thing breaks.

Human / LLM Symmetry
Either side composes, reviews, modifies
Human Audit
Every decision traceable
LLM-Safe Creation
Constrained error space
Security
Per-node access control. Multi-tenant isolation.
— Click a block to explore

The Architecture

Each requirement builds on the ones below. Security is the foundation — without it, nothing is trustworthy. LLM-safe creation and human audit sit in the middle — both require secure infrastructure. At the apex, human/LLM symmetry requires all three: safe creation, auditability, and enforced boundaries.

The Spec

Five components.

The irreducible core. Remove any one and at least one requirement fails. Click a component to see what breaks.

node
Minimal compute unit with declared dependencies.
LLM returns to writing arbitrary code. Error space is unbounded.
LLM-Safe Creation
edge
Dependency as explicit data — the links between nodes.
Dependencies are implicit. Causal chain breaks.
Human Audit
propagation
Change → downstream notified instantly, no stale state.
Manual update coordination. Back to polling, back to vibe coding.
LLM-Safe CreationHuman Audit
describe()
Structure visible, persistable, diffable at any moment.
Structure is invisible. Not persistable, not diffable.
Human AuditSecurity
GraphSpec
Declarative, LLM-readable format for compose / review / modify.
LLM can't structurally operate on the graph.
Human / LLM Symmetry
— Ship It

It exists.

One protocol. Two runtimes. TypeScript and Python, same spec, same guarantees. Zero dependencies. Tree-shakeable. One primitive.

TypeScript
$npm i @graphrefly/graphrefly
View on GitHub ↗
Python
$pip install graphrefly
View on GitHub ↗
Copied to clipboard