Private Intelligence Platform — Developer

Flows that are deterministic.Steps that aren't.

LLMs are non-deterministic. That's fine — that's their strength. But they cannot be trusted to orchestrate. reBe gives you both: a WASM flow runtime that is deterministic by construction, and LLM steps that operate within it under contract.

The flow runs on your hardware, reads from your corpus, executes locally, and writes to an immutable chronicle. No cloud dependency. No usage meter. No data leaving your realm.

The Problem

Agents don't scale. Flows do.

Current approach

LLM as orchestrator

  • Agent decides what to do next — non-deterministic by definition
  • No audit trail — "the model chose this path" is not compliance
  • Scale by paying more tokens — the meter never stops
  • Data sent to the cloud to be understood, then acted on
  • Fails silently — hallucinated steps look identical to real ones
reBe approach

WASM flow as orchestrator

  • Flow structure is deterministic — LLMs fill bounded steps within it
  • Chronicle records every step: actor, input, output, duration
  • Scale by adding hardware — no per-execution cost
  • Corpus stays on your device — LLM reads locally, acts locally
  • Every step has a schema — output is validated before the flow continues
How it works

Deterministic structure. LLM steps. Local execution.

The flow is a contract. Each step has a type, an input schema, and an output schema. LLM steps are just another step type — bounded, testable, auditable.

Flow: monthly-spend-review
1corpus.querydeterministic

your email, last 30 days

2llm.categorisellm step

local model, bound schema

3llm.summarisellm step

local model, bullet format

4chronicle.appenddeterministic

immutable audit record

Structure is deterministic. LLM steps are bounded. Every run produces a chronicle entry. Auditable by design.
flow-definition.ts
import { getTheCy } from '@thecy/realm-client'

const cy = await getTheCy()

// Define a flow from composable WASM steps
const flow = cy.flows.define({
  id: 'monthly-spend-review',
  steps: [
    { id: 'collect',  type: 'corpus.query',   params: { slice: 'finance', window: '30d' } },
    { id: 'classify', type: 'llm.categorise', params: { model: 'local', schema: SpendSchema } },
    { id: 'summarise',type: 'llm.summarise',  params: { model: 'local', format: 'bullets' } },
    { id: 'record',   type: 'chronicle.append', params: { thing: cy.identity.thingId } },
  ],
})

// Run it — deterministic structure, LLM fills 'classify' and 'summarise'
const result = await flow.run()
// → chronicle entry written, audit-safe, ran entirely on your hardware
The boundary

Where determinism ends. Where trust begins.

You don't choose between deterministic and non-deterministic. You compose them. The flow enforces the structure. The LLM fills in the creative work. The chronicle makes it auditable either way.

Aspectdeterministicllm step
OrchestrationWASM flow definition — same structure every runLLM generates the flow from a natural language description
Step executionPure functions: input schema → output schema, no side effectsLLM fills a step: prompted with context, bound by output schema
AuditChronicle records every step: inputs, outputs, duration, actorLLM response is recorded verbatim — non-determinism is traceable
TestingFlow structure is testable — run against fixture corpus, assert outputsLLM steps are evaluated — prompt + rubric, scored by local model
Trust boundaryCompliance-safe: auditor sees the flow, the inputs, the outputsLLM steps are bounded — they cannot escape the flow contract
Architecture

Four tiers. One flow definition.

A flow written at T0 runs identically at T1, T2, and T3. Scale is a deployment decision, not a rewrite. The WASM binary is the same — only the compute and corpus change.

T0
Browser
Pure WASM. Zero server.
Flows run entirely in the browser tab. No network. No API key. OPFS for state.
WebLLM inference — distilled models, downloaded once, run locally forever.
  • WASM flow runtime — deterministic, sandboxed
  • LLM step execution via WebLLM (no cloud)
  • OPFS encrypted local storage
  • SQLite-vec semantic search
  • BroadcastChannel IPC across same-origin pages
T1
Node
Go binary. Any LAN device.
Flows coordinate across the user's devices. Node is the corpus broker and LLM router.
Model runner on local hardware. Cost: electricity. Privacy: absolute.
  • Flow coordinator — fan-out across node lattice
  • IMAP/JMAP corpus sync (email, docs, calendar)
  • FTS5 + semantic search across slices
  • Model runner routing — model selection per step
  • Runs on Mac, Linux, Raspberry Pi
T2
Realm
K3s + wasmCloud + NATS.
Multi-user flows. Corpus federation. Audit-grade chronicle for every execution.
Lattice-aware routing — best available model across the realm, cost-optimised.
  • wasmCloud actor deployment for flow components
  • NATS lattice — invoke any node in the realm
  • Multi-device, multi-user corpus federation
  • Trust ledger — immutable execution chronicle
  • No hyperscaler dependency

T3 — CSP Edge available for operator deployments. thecy.io/operators ↗

Philosophy

Every human becomes a programmer.

Semantic interfaces, not syntax

You describe a process in natural language or voice. A local distilled LLM generates a verifiable flow from composable WASM parts. You inspect it in plain language — “this step looks wrong” — and regenerate. No IDE. No deployment pipeline. No syntax error.

Creator and destroyer of all things digital

The flows you define are yours. They live in your realm. They read from your corpus. They write to your chronicle. You grant permission for other Things — people, agents, automations — to invoke them. You revoke it. You delete them. No platform owns the execution.

Composable parts, not monolithic apps

Every step is a WASM module. Modules are composable, testable in isolation, replaceable without touching the flow. The LLM-generated step for “classify by project” can be replaced with a deterministic classifier once you have enough examples. The flow contract stays the same.

Auto-generation and auto-testing

When a local LLM generates a flow, it also generates tests — prompt templates, fixture corpus, expected output schemas. Before the flow is deployed, it runs against your real data. You see pass/fail per step. If a step fails, you tell the model what was wrong in plain language and it regenerates.

Function-specific, distilled, local

Not one large general model. A portfolio of small, distilled models optimised for specific functions: classification, summarisation, extraction, code generation, question-answering. Each runs locally at T1. Each is replaceable. None requires a cloud API call.

Trusted by construction, not by policy

A compliance team can read the chronicle entry for any flow execution: who invoked it, what corpus slice was read, what the LLM step received as input, what it returned, what happened next. The trust is in the structure, not in a vendor's audit dashboard.

vs. workflow tools

Not a workflow engine. A flow runtime you own.

Zapier, n8n, Temporal, Prefect — all SaaS platforms or self-hosted services. Your flows depend on their uptime, their pricing, their architecture decisions. reBe flows are WASM binaries. They run wherever WASM runs.

DimensionWorkflow toolsreBe flows
OrchestrationCentral SaaS schedulerWASM flow — runs anywhere
LLM roleAgent drives the processLLM fills bounded steps
DataSent to cloud for processingStays on device/LAN
Audit trailPlatform logs (opaque)Chronicle — cryptographic, immutable
ScalePay per execution, per seatHardware you own — no meter
DeploymentSaaS or self-hosted serviceWASM binary — browser, Pi, K3s, edge
IdentityOAuth, platform-managedCryptographic realm identity
reBe Developer

This is already built. Open it.

The Flow Builder, Developer IDE, and reBe Agent are running inside reBe right now. Your realm boots in the browser — no install, no account, no cloud.

reBe Flow — monthly-spend-review
XyFlow canvas
corpus.queryfinance · last 30 daysdetllm.categoriselocal model · SpendSchemallm stepllm.summariselocal model · bulletsllm stepchronicle.appendimmutable audit entrydetNODE TYPEStriggeraicorpus-readcorpus-writehttpconditiontransformnotify
deterministicllm step11 node types · XyFlow canvas
Flow Builder

Build a flow

Open the XyFlow canvas. Drag nodes, wire steps, run your first flow against real corpus data. Deterministic by construction.

11 node typesLLM stepsrun historytemplates
Developer IDE

Open the IDE

PTY terminal, flow orchestrator, provisioning wizard, party line. Everything you need to build and operate a sovereign app.

PTY shellorchestratorprovisioningparty line
reBe Agent

Ask reBe

Describe what you want to automate. reBe generates a flow from your corpus context. Natural language to WASM flow.

corpus-awarelocal LLMflow generationvoice

Boots in the browser. No install. Your realm, your data, your hardware.

Get started

Build the first sovereign flow.

realm-client is in active development. The T0 browser layer is production-ready. T1 node and T2 lattice federation are in beta. We're working with a small group of developers building the first sovereign-native applications — flows that run on hardware you own, read from corpus you control, and write to an audit log that belongs to you.

theCy is the platform. reBe is the personal layer built on it.

IT / Partners — Deploy and Operate

Deploy sovereignty as a managed service.

For IT teams, managed service providers, and VARs — reBe is a deployable sovereign stack. You provision realms for clients, manage the fleet, brand the experience, and own the relationship. Your clients get sovereign AI and data sovereignty. You get recurring infrastructure revenue without hyperscaler margin compression.

White-Label

Deploy Under Your Brand

Deploy reBe under your brand. Your clients see your name on their sovereign platform. You manage the infrastructure, they own their data. The relationship is yours.

Fleet Management

Central Control

Provision, update, and manage hundreds of client realms from one console. Per-realm policy. Per-realm corpus governance. Central audit trail.

Specialist Corpus

Vertical Expertise

Build and ship pre-loaded corpus for a vertical — healthcare, legal, construction, finance. Your expertise + reBe sovereignty = a differentiated offering.

Partner Economics

MSP Pricing

No per-token cloud cost means your margin is your infrastructure margin. Fixed hardware cost, unlimited inference — unit economics improve with scale.

Become a reBe partner.

Deploy reBe as a managed service. White-label it. Build vertical-specific corpus. Earn recurring infrastructure revenue. Your clients get sovereignty. You own the relationship.

Contact us →
Open in Studio

The Architecture of Sovereignty: Compiling Concepts into Reality

Open this Studio piece and ask reBe how to build a capability — Things, asCode, Operator.

Open the piece on Studio →