How Humonix is built.
One OS, a Brain per business.

For engineers and investors: the layers behind Humonix, the stack each one runs on, and the rules that keep AI agents safe inside a business.

CHANNELSWeb chatTextEmailPhoneMCPHUE · THE INTERFACEONE THREAD PER PERSON · PERMISSION-SCOPEDALLOMETRY · THE DECISION OSData spinePostgres + RLSAgentsbackground jobsApprovalsa person decidesLedgerhash-chainedYOUR SYSTEMSQuickBooksHubSpotShopifyStripeMODEL GATEWAYTASK → TIER → PROVIDER → FALLBACKClaudeGPTGeminiOpen-weight

How a request moves

ILLUSTRATIVE TRACE, NOT REAL OUTPUT
thread.append channel=web person=owner
records.read scope=person rows=3
connector.read source=quickbooks invoices=2
gateway.route task=draft tier=standard → claude
draft.hold reason=changes_price waits=approval
ledger.append prev=7f3a… hash=c91e…

The layers

Allometry: the Decision OS
Today the client app, the API, the MCP server and A2A run in one Next.js 16 app on SQLite (Fly.io, Toronto, replicated by Litestream): every query is filtered by the company in the session, with our own sessions and API keys, connectors to the tools a business already uses, and a ledger of every action. Allometry is the planned data layer: a Postgres spine (Drizzle ORM, row-level security per tenant) and background agents; it is not in production yet.
Draftprev 0000…hash 7f3a…Approvedprev 7f3a…hash c91e…Price setprev c91e…hash 2b80…Sentprev 2b80…hash e46d…LEDGER · EACH EVENT CARRIES THE HASH BEFORE ITVERIFIED WITH NO DATABASE AND NO KEYILLUSTRATIVE HASHES
The model gateway
Every model call goes through one gateway: a task picks a capability tier (fast, standard, deep), the tier picks an ordered list of providers, and the call falls through to the next one if a provider fails. Residency rules are enforced at the gateway, so a tenant pinned to a local model never reaches a hosted one. Claude by default; OpenAI, Google and open-weight models behind the same interface.
TASKTIERPROVIDERS, IN ORDERDraft a quoteFASTSTANDARDDEEP1 Claude2 GPT3 GeminiPinned tenantLOCAL ONLYLocal modelHOSTED: BLOCKED HERERESIDENCY IS ENFORCED AT THE GATEWAY
The harness
Each agent gets least-privilege access to the tools its job needs. Anything that spends money, changes a price or goes public stops at an approval. Content from emails, pages and documents is treated as data, never as instructions. Every tool run writes the same ledger events whether it came from the console, the API or MCP.
AgentRead recordsDraftChange a priceSend an emailMove money
QuotesYesYesApprovalApprovalNo
RepliesYesYesNoApprovalNo
ReordersYesYesNoApprovalApproval
CollectionsYesYesNoApprovalNo

Example policy Quotes → Change a price: the agent drafts, a person decides.

The door: MCP and API
Allometry exposes its tools over MCP (JSON-RPC 2.0) and a REST API built on the same cores. Each agent has its own token and budget, so Hue, Claude or a client’s own agent all go through one governed door.
HueClaudeClient’s agentMCP · RESTTHE SAME CORESthe same ledger events429 · OVER ITS BUDGET
Hue: the interface
Web chat is live; text, phone, email, and Claude and ChatGPT over MCP are in early access. One thread per person. A signed-in person only gets answers from records they may see, each with its source. Read-only: changes are drafted for approval.
HQ: our own operations
Humonix runs on its own platform. HQ is a Next.js app with SQLite on a Fly.io volume, continuously replicated by Litestream, that handles intake, CRM, outbound, invoicing and approvals for every brand in the group.

How we build

Rules first, models where they’re measured
Deterministic rules decide the unambiguous cases; a model handles the rest with a confidence floor; every model or prompt change is scored on a fixed eval set before it ships. Example: how we sort replies to our own outreach.
Drafts, not writes
Agents propose; people approve. No agent writes to a system of record without an approval.
Pure, tested modules
Decisions live in pure functions with unit tests (node:test, Vitest); I/O stays at the edges.
Bilingual from day one
English and Québec French in every interface, email and document.
Built with Claude Code
We work alongside Claude Code with project rules, skills and hooks; every change runs against tests before it ships.
Hosted close to home
Client data is hosted close to home, in North America: Toronto by default, with US hosting for US clients on request (their own instance, in Virginia). Canadian clients’ data is hosted in Canada (Québec Law 25).

The stack

LayerStackHosting
humonix.com siteVite, React, prerendered pages, English and FrenchFly.io, Toronto
HQNext.js 16, Node, SQLite + LitestreamFly.io, Toronto
AllometryNext.js 16, TypeScript, Postgres + Drizzle, Inngest, Clerk, WorkOS, SentryNot yet in production
ModelsAnthropic Claude (default), OpenAI, Google, open-weight models, TypeSafe Jev for fast typed decisionsProviders, or local for region-pinned tenants
InterfacesMCP (JSON-RPC 2.0), REST API, webhooks

Questions

Is the Brain a data lake?

Not in the usual sense. A data lake stores raw data for analysts. The Brain keeps the data agents need (customers, products, prices, documents) with the business’s know-how and every approved decision, in one shared data model, and connects to the systems where the rest lives. The business owns it and can export it any time.

What do you mean by harness?

The layer around AI agents that makes them safe in a business: it gives them data and knowledge, limits what each can access, decides what needs an approval, and logs every action.

Can a client bring their own model keys?

Yes. They then pay the provider directly, and the gateway uses their keys. See the platform fee page.

Where is data hosted?

In North America, close to you: Toronto by default, with your own US instance on request. It is never used to train shared models.

Do you use MCP?

Yes, both ways: Allometry exposes its tools over MCP, and Hue answers inside Claude and ChatGPT through a public, read-only connector, in early access.

Build it with us

An early-stage, founder-led company building an operating system for SMBs, in English and French. If this is the work you want to do, say hello, even if the role isn’t listed yet.

Talk to us