How we build it

Envoy, and the ideas it's built on

This is the system we test our own ideas in: Envoy, the assistant layer, and Envoy Warehouse, the governed data layer underneath it. Here's how each piece works and why we made it that way.

The idea

An ontology is just a clear model of how a world fits together.

A general model is fluent but rootless — it doesn't know a given organisation's terms, its systems, or what its numbers actually mean, so it improvises. The bet we're testing is that giving it roots — an explicit model of the domain to answer from — beats prompting harder. Everything below is what that looks like when you build it properly.

The platform

Two halves, one foundation

The surface people talk to, and the governed data layer it stands on — joined by a shared model of the domain.

The surface · Envoy

The layer people actually talk to

A conversational workspace with documents, agents, and skills — the place where the ontology has to earn its keep in ordinary use.

Read on →
⌖ Your ontology — the shared model

The foundation · Envoy Warehouse

The governed data layer underneath

Data modelled as typed objects and named metrics computed deterministically — served to any AI through a locked-down MCP server. Trustworthy numbers, never raw rows.

Read on →
Envoy · the assistant layer

AI that speaks the local language

Every organisation has its own vocabulary, and generic assistants flatten it. So Envoy carries the domain's terms into retrieval, into the agents, and into how answers get cited. The design rule we hold ourselves to: it should make people better at their own work, not stand in for them.

How the skills work
Workspace · Finance
How did sales track last week versus the same week last year?
Sales were up 8.4% week-on-week against last year, with gross margin holding steady. ↳ from query_metric(activity, average · segment) · ~134 tokens · cited

Conversational workspace

A clean chat interface your team already knows how to use — no training required.

Your documents, retrieved

Upload to a workspace; Envoy retrieves and cites the right passages with hybrid search (RAG).

Agents, tools & skills

Ready-made agents out of the box, plus bespoke ones tuned to your terminology and workflows.

Workspaces

Spaces tailored to each team — each with its own documents, agents, and context.

Deep integrations

Connects to the systems you already run — CRM, accounting, storage, comms, data — via a real integration framework.

Admin console

Your team manages users, branding, agents, tools, and usage themselves. No black box.

Specialist agents

Agents built for the actual job

Not one generic chatbot, but focused agents — each tuned to a real task and powered by composable, portable skills. These are the ones running today; the list below is a fair picture of what the pattern is good at, and where it needs a specialist to sit beside it.

Legal · construction

Contract Review

Compares an AS 4902 contract against your organisation's preferred positions clause-by-clause, flags departures from your playbook and the standard's defaults, and renders a departures table right in the chat.

Bids & tenders

Tender Response

Draws on your past tender responses to draft consistent, well-supported answers — a single focused question, or a whole tender at once — ready to review and paste.

Commercial Contract Review

Legal

Reviews vendor agreements, MSAs, and SaaS subscriptions against your Legal Playbook — a deviation memo with severity, redlines, and approval routing.

Privacy / DPA Review

Legal

Reviews Data Processing Agreements term-by-term; detects processor vs controller direction and flags transfer mechanisms and sectoral overlays.

Employment Review

Legal

Reviews offer letters, employment and contractor agreements against your Legal Playbook, with jurisdiction-specific research.

NDA Triage

Legal

Fast GREEN / YELLOW / RED triage of inbound NDAs — so only the ones that genuinely need lawyer time get routed there.

IP Clause Review

Legal

Reviews IP clauses — assignment, licence, warranties, indemnities — against your playbook, with assignment-gap detection.

Accounting Review

Finance

Reviews contracts for accounting implications — revenue recognition, leases, modifications — under IFRS, US GAAP, or AASB.

Research Analyst

General

Combines web research with your workspace documents to produce sourced, structured briefings.

Summariser

General

Short, medium, or detailed summaries of any document you upload.

Curious how one of these is put together? Ask and we'll walk you through it →

Skills

Composable know-how, yours to keep

A skill is a small, open SKILL.md file that teaches an agent one job — when to step in, how to do it, which tools to use, plus any reference files it needs. Agents load them only when relevant. Some run on demand; others are always on, encoding conventions like citation discipline and locale.

legal-playbook-setup On demand

An interactive interview that captures your team's positions on NDA terms, vendor contracts, escalation, and delegations.

nda-review On demand

Fast GREEN / YELLOW / RED triage of inbound NDAs, so legal time only goes where it's genuinely needed.

extract-tender-questions On demand

Parses a tender RFT into a structured list of the questions and criteria a bidder must answer.

standards-reference On demand

Cites and applies external standards correctly — IFRS, US GAAP, AASB, and data-protection regimes.

research-brief On demand

Produces a structured research brief: background, key findings, sources, and open questions.

document-extraction-helper On demand

Pulls structured data — tables, key/value pairs, entities — out of PDFs, CSVs, and spreadsheets.

source-fidelity Always on

Strict citation discipline: quote the source verbatim, then paraphrase. Never invent a figure.

aussie-english Always on

Locks the conversation to Australian English spelling and conventions.

user-locale Always on

Carries the user's locale so numbers, dates, and currency are formatted consistently.

nda-review/SKILL.md
---
name: nda-review
description: Triage inbound NDAs into GREEN / YELLOW / RED.
when_to_use: an NDA is attached for review.
license: your-organisation
allowed-tools: document_search, read_attachments
user_invocable: true
---
# Instructions + reference files + assets travel together.
# Open format → runs in Claude (Code & API) and other tools.

Export — take it anywhere

Download any skill as a name.zip — its SKILL.md plus every reference file and asset. Because it's the same open Agent Skills format Claude uses, it runs in Claude (Code & API) and other agent tools, not just Envoy.

Import — bring your own

Add a skill by pasting its SKILL.md, uploading a .md or .zip, or filling in a guided form. Envoy lints it, versions it, and tracks every change — so your team can author, fork, and refine skills with no lock-in.

MCP · query_metric
# Ask for typed objects + named metrics — not raw rows
query_metric(object: activity,
metrics: [total, count, average],
dimensions: [segment])
# → deterministic, computed in Postgres, ~134 tokens
segment‑a  total 000,000  count 0,000  average 00.00
segment‑b  total 000,000  count 0,000  average 00.00
segment‑c  total 000,000  count 0,000  average 00.00
Envoy Warehouse · the data layer

Numbers AI can actually trust

Envoy Warehouse ingests your data and models it as an ontology — typed objects, named metrics, and dimensions. Metrics are computed deterministically in the database and served to AI through a locked-down MCP server, so models reason over governed facts instead of guessing from raw tables.

  • Deterministic metrics. Every number is computed in Postgres from a defined formula — same question, same answer, every time.
  • Token-efficient by design. A full management summary is ~134 tokens, not thousands of rows — fast, cheap, and within context.
  • Locked-down MCP server. Nothing is discoverable without a credential. Connect Claude (Desktop, web, Enterprise) via token, OAuth, or Microsoft Entra SSO.
  • Configuration, not rewrites. Onboarding a new client or domain is a new ontology config — the generic core never names a client.

A design principle

What should be specific, and what shouldn't

Vocabulary, workspaces and skills belong to the domain and should be configurable. The engine underneath shouldn't know a single customer's name — the moment it does, every fix becomes a fork.

Vocabulary is configuration

Terms, documents and working conventions live in the ontology config and the workspace, not in the code — so they can change without a release.

Narrow beats general

A tightly scoped agent that knows one job and one vocabulary reliably outperforms a general assistant given the same task. This is the clearest pattern we've found so far.

No lock-in

Skills that travel with you

Skills are open SKILL.md files — the same format Claude uses. Export them and they run in Claude (Code & API) and other tools, not just Envoy. Know-how someone wrote should outlive the tool it was written in.

See how export & import work →

Trust & security

Your brand, your data, honest answers

Security is structural — enforced by the database and the protocol, not by hopeful application code. And because the numbers come from the warehouse, the AI can't make them up.

Your brand and domain

Runs white-label under your logo, colours, and domain. Your team and customers see your product, not ours.

Schema-per-tenant isolation

Each tenant's data lives in its own PostgreSQL schema — isolation enforced by the database itself.

Encrypted credentials

Integration credentials are encrypted per tenant; authorisation is explicit on every action.

Deterministic, cited answers

Numbers are computed in the warehouse and cited. Same question, same answer — no hallucinated metrics.

Locked-down MCP

Nothing is discoverable without a credential — token, OAuth, or Microsoft Entra SSO. Every call is scoped.

Full observability

Every agent run, tool call, and cost is captured — so you always know what your AI is doing.

The method

The order we've learned to do it in

Modelling before building, and handing over the controls before anyone gets attached to us holding them. Take it and use it — the sequence matters more than the software.

01

Model it

Learn the problem and the words people already use for it, then write that down as an ontology. Skipping this step is where most AI projects quietly go wrong.

02

Wire it

Set up the agents, metrics, tools and integrations against that model — configuration, not a new codebase each time.

03

Let go

Hand the controls to the people doing the work. If they can't change it without us, we built it wrong.

The Ontology AI

Take the ideas, and tell us where we're wrong

None of this is settled. If you're building something in the same territory, or you think a decision here is a mistake, we'd genuinely like to hear it. We keep publishing what we learn either way.