The harness

How Humonix
runs agents.

The same tools behind every door, the right Claude model for each task, your data hosted close to home, in North America, and a person’s yes before anything spends money or goes public. Written for owners, and for the IT person they ask.

This page describes what runs today. Anything not built yet is marked “coming”.

01

One tool registry. Four doors.

Every capability Hue has is declared once, in one registry. Hue uses it in the app and on your site; your own AI assistant reaches the same tools over MCP; your code over the REST API; another company’s agent over A2A. One set of tools means one set of rules: a check added once holds at every door.

Reads change nothing. A proposal adds a draft to the queue a person already approves. No tool sends, pays, publishes or deletes on its own.

  • You and your customers

    Hue

    In the app and on your website.

    Live
  • Your AI assistant

    MCP

    Claude, ChatGPT, Cursor: OAuth 2.1, scopes, one company per connection.

    Live
  • Your code

    REST API

    A company API key, OpenAPI 3.1, a public catalog.

    Live
  • Other agents

    A2A

    An Agent Card; tasks come back with a result.

    Live

The one door every call goes through

Whichever door a call comes in by, it passes the same checks, in this order.

  1. 01The tool exists on that door
  2. 02The caller’s scope covers it
  3. 03Your company only, never one named in the request
  4. 04Logged: who asked, and from where
  5. 05The person’s role allows it
  6. 06Its own checks, then a rate limit
  7. 07It runs: a read, or a draft for approval
  8. 08Metered: tool, door, outcome, time
The same registry generates Hue’s tools, the MCP server and the REST API.

02

A model router, not a model.

Callers name a kind of task, never a model. The router picks the tier for that task, the Claude model for the tier and the key (ours or yours), and checks the month’s limit before the call goes out. Every call lands on a usage ledger at the provider’s list price.

Claude is the default and, today, the only model Hue works with. Other providers come in only to check how their AI answers name you.

TaskTierModelKeyLimit

Most capable

Claude Opus 5.5

Plans, goals, strategy, data structure

BalancedDefault

Claude Sonnet 5.5

Drafts, writing, answers to customers, chat, research, using tools, facts about your business

Fast

Claude Haiku 5.5

Sorting messages, routing, pulling details out, short summaries

When a model is overloaded or down

  1. Retry, with a short wait
  2. Then the next tier down: an answer from a faster model beats no answer
  3. Planning steps up a tier instead, never down
  4. A model that keeps failing rests for 30 seconds
  5. The provider never changes silently

Providers

  • Claude (Anthropic) Live

    The default, for every task Hue does.

  • OpenAI, Gemini, Perplexity Live

    Asked the questions your buyers ask, to see where AI answers name you (AI visibility). Logged at their list price.

  • Other providers for Hue’s work Coming

    Routing a task to GPT or Gemini, with a fallback across providers.

  • Open-weight models Coming

    For bulk work or a model you host. When they pay off →

Model names as of November 2026; a newer model replaces one per tier without moving your data.

03

Your settings, in the app.

Owners and admins steer the router from Settings → Hue’s models: the tier for each kind of task, what happens past the included volume, a monthly ceiling on AI cost, and their own Anthropic key. The included volume itself is set by your plan.

A tier above a task’s default needs your own key, since your account pays for it. On the free plan Hue always stops at its volume.

Settings → Hue’s models

Your plan includes a set number of answers a month; this card shows how many are used so far.

Past the included volume

Monthly limit on AI cost (optional)

150USD

At this amount Hue stops for the month, whatever you chose above.

Tier for each kind of task

  • Answers to customers
  • Drafts
  • Sorting messages
  • Plans

Your own Anthropic key

Key ending 7Q2x · added 2026-11-02

Sealed when you add it and never shown again: only the last 4 characters. On your key, your Anthropic account pays and we never cap or bill Hue for it.

Sample values; the wording is the app’s. Nothing here is saved.

04

Your data, hosted close to home.

Client workspaces and our own systems run on Fly.io machines in Toronto, Canada, on encrypted volumes with continuous backups and daily snapshots. A US client can ask for its own instance in Virginia for its synced records and search index; Canadian clients’ data is hosted in Canada. Your data is never used to train shared models: Claude is called through Anthropic’s API, under business terms that exclude training.

Export everything, in open formats, any time. When you leave, your data is deleted 30 days after you confirm.

Toronto, CanadaFly.io · yyz

Workspaces, records and the ledger
Hosted in TorontoLive
Training shared models on your data
NeverLive
Export
Everything, open formats, any timeLive
When you leave
Deleted 30 days after you confirmLive
Your own provider key
Anthropic, sealed, last 4 shownLive
Model calls
Anthropic’s API, no trainingLive
Claude on AWS Bedrock in Canada
For residency of the model calls tooComing
Your own instance or white label
For EnterpriseComing
“Coming” means designed, not built: we say so until it runs.
  • Québec Law 25

    Marketing consent recorded separately, with its wording and time. Hue says it is an AI. Access and deletion requests answered within 30 days.

  • PIPEDA

    Personal information is processed only to run the service, under our data processing terms, with every subprocessor listed publicly.

  • CCPA

    We never sell your visitors or leads, and never use them for our own marketing.

Notes, not legal advice. SOC 2: not certified yet; the security page says where we are.

05

Nothing spends or goes public without a yes.

Agents draft; people decide. Anything that spends money, changes a price, sends or goes public waits in Approvals for a named person. Try it: approve or skip a card below.

Every action, outcome and approval goes on your company’s ledger with who and when.

HumonixACME Field ServicesSample data
Approvals: nothing goes out without a yes. Try Approve or Skip; every decision goes on the ledger. Sample data · ACME Field Services, an invented company.

The ledger, hash-chained

Each entry carries the hash of the one before it. Change or remove an entry and every hash after it stops matching, so anyone can check that nothing was edited afterwards. You can read it and export it; nobody rewrites it.

  1. #1041 · 08:42Action

    Drafted quote Q-2026-0157 for Lakeview Hotel

    Run module agent

    prev 9f3c…a1e0hash 4be0…7d52

  2. #1042 · 08:51Approval

    Approved quote Q-2026-0157

    Dana Ruiz, owner (sample)

    prev 4be0…7d52hash c27a…1913

  3. #1043 · 08:52Outcome

    Quote sent from your mailbox

    Run module agent

    prev c27a…1913hash 06df…e2b8

Sample entries for ACME Field Services, an invented company. Each hash is SHA-256 of the previous hash and the entry.

06

Are you ready for agents?

An agent works only from what your business has written down and what it can reach. Here are six of the twelve things the agent readiness audit checks. Tick what you have today.

The full audit is free, needs no sign-up, scores all twelve and tells you what to fix first.

0of 6 in place. A quick read, not the audit’s score.

Take the twelve-question audit
Nothing you tick leaves this page.

Questions IT people ask

Which AI models does Humonix use?

Claude by default, through Anthropic’s API: Claude Opus 5.5 for plans and strategy, Claude Sonnet 5.5 for drafts and answers, Claude Haiku 5.5 for sorting and routing. OpenAI, Gemini and Perplexity are asked your buyers’ questions to check AI visibility. Routing Hue’s own work to other providers, and open-weight models, are coming.

Where is our data hosted, and is it used to train AI?

In North America, close to you: client workspaces run on Fly.io in Toronto, Canada, like our own systems, and a US client can ask for its own instance in Virginia. Your data is never used to train shared models; Claude is called under Anthropic’s business terms, which exclude training. Claude on AWS Bedrock in Canada, so the model calls stay in Canada too, is coming.

Can we use our own Anthropic key?

Yes. An owner or admin adds it in Settings → Hue’s models. It is sealed when stored and only its last 4 characters are ever shown. On your key, your Anthropic account pays and Humonix never caps or bills Hue for it.

Can an agent send, spend or publish without us?

No. Agents draft; anything that spends money, changes a price, sends or goes public waits for a named person’s approval, and the decision is written to a hash-chained ledger with who and when. Over MCP, REST and A2A there is no tool that sends, pays, publishes or deletes.

What happens to our data if we leave?

You can export everything in open formats at any time. When you close your account, your data is deleted 30 days after you confirm.

Start with
twelve questions.

The agent readiness audit tells you what to fix before an agent touches your work. Free, no sign-up, and your answers stay in your browser unless you send them.

Read next

Every Desk (Attract, Create, Inbound, Outbound, Quote, Collect, Nurture, Grow, Build, Optimize, Measure, Research) and Know run on this harness.