The Naevum stack

Own the system your organization runs on.

Most organizations rent their memory, their workflows and their AI from several vendors at once. The Naevum stack puts them in one place that you control, and you can start with nothing more than your folders.

Naevum Labs owns and develops the entire stack. Metahive is its research, compute and Tier 1 line.

Plan your migration

The stack at a glance.

Three tiers on shared infrastructure and research. Metahive carries research, compute and Tier 1; Naevum Labs builds Tiers 2 and 3.

  1. Tier 3 · InstitutionalNaevum Labs
  2. Tier 2 · VerticalNaevum Labs

  3. Metahive research, compute and Tier 1

    1. Tier 1 · Included
      • Metahive
      • Folder ingestion
      • Knowledge base and memory
      • Program proposals
      • Governed MCP gateway
      • Evidence record
    2. Infrastructure
    3. Research
      • Research track
      • Evidence grading

Foundations under all tiersAgency Engine · Black Box Factory · Conductr

What the research says

What the rented stack costs before anyone does any work.

Three figures from the suites most organizations already pay for.

$9,4551
Median SaaS spend per employee in 2026 (vendor dataset, skewed to large companies)
36%2
Average share of SaaS licenses left unused, against the vendor’s utilization levels
$7.003
Microsoft 365 list prices from 1 July 2026: most business and enterprise plans rose

An organization pays for its software per employee, every year, including the seats nobody opens. Each seat rents a piece of the organization’s memory and workflow back to it. The Naevum stack starts from the folders the organization already holds and connects to the rented tools, so a team can see which seats the work still needs before it retires any.

01 / 03 · Tier 1 · Included

Share a folder. Get a working environment.

Delivered by Metahive, the research, compute and Tier 1 line of Naevum Labs.

Point the system at the folders your organization already has. It reads them, records where every fact came from, and builds a knowledge base and a memory you can question. Then it proposes the programs, skills and workflows your work needs, and you approve them.

  • Folder ingestion

    Reads the text files in the folders you share (Markdown, plain text, JSON, CSV and YAML) and records the path and date of each one.

  • PDF and Office files

    PDF, Word, Excel and PowerPoint files are listed as gaps today. Their ingestion is being built.

  • Knowledge base and memory

    Built from the index, so anyone in the organization can ask a question and follow the answer back to its files.

  • Program proposals for review

    The system proposes programs, skills and workflows for your work. A person reviews each one, and none is installed without approval.

  • One governed MCP gateway

    Works through one governed gateway with the command-line coding agents your team already uses: Claude Code, Codex, Gemini CLI and OpenCode.

  • Claude app, ChatGPT app and VS Code connectors

    Connectors for the desktop and editor tools are being built on the same gateway and the same rules.

  • The evidence record

    Every consequential action lands in a hash-chained, Merkle-anchored, timestamped record you can verify, which shows the history was not rewritten. Built for the AI Act’s obligations: roles allocated, mechanisms built, evidence kept.

How does a firm move onto its own stack?

A managing partner decides what to share and approves every program; the stack does the rest and records it. Illustrative, with a synthetic 40-person engineering firm.

  1. 1The managing partner points the stack at the folder tree the firm already has: projects, templates and policies.Decision: Which folders to share, and which to leave out.
  2. 2Each text file is indexed with its path and date, so every fact can be traced to its file. PDF and Office ingestion is in development.
  3. 3A knowledge base and a memory are built from the index. Anyone at the firm can ask a question and follow the answer back to its files.
  4. 4The stack proposes programs, skills and workflows for the firm’s work. The managing partner reviews each one, and nothing is installed without that approval.Decision: Approve, change or reject each proposed program.
  5. 5The engineers’ coding agents connect through one governed gateway. Connectors for the Claude app, the ChatGPT app and VS Code are in development.
  6. 6Every consequential action, from indexing to an approved program’s run, is written to a hash-chained, timestamped record the firm can verify.
Select any part of the drawingScroll to step · Esc to skipFig. 2
  1. 1An engineer saves a revised client specification into the shared folder. The stack indexes the change with its path and date.
  2. 2The engineer asks her coding agent what changed in the client’s requirements. The answer comes from the firm’s knowledge base, with the files it used.
  3. 3The design-review program the partner approved in the first days runs on the change and prepares a draft review.
  4. 4The draft, the files it relied on and the program that produced it are written to the record. The engineer decides what happens to the draft.Decision: Send, edit or discard the draft.
Select any part of the drawingScroll to step · Esc to skipFig. 2
  1. 1An auditor asks what the firm’s AI agents did with one client’s files during a given week.
  2. 2The team queries the record for that client and that week. Each entry shows what was done, by which agent or program, and when.
  3. 3A Merkle inclusion proof shows that each entry belongs to the record, and a timestamp fixes when it was written.
  4. 4The managing partner decides what to disclose. The auditor checks the proofs with standard tools and sees that the history was not rewritten.Decision: What to disclose to the auditor.
Select any part of the drawingScroll to step · Esc to skipFig. 2

02 / 03 · Tier 2 · Vertical

Your industry, as an operating system.

Built and delivered by Naevum Labs.

A vertical adds what general tools leave out: the rules of your jurisdiction, expert agents for your functions, and a private boundary for model work. It is assembled from bio-inspired systems, each with one job, and installed as a cluster or one at a time.

  • Local legal frameworks

    The rules of your jurisdiction and how they are read, as a source the vertical’s agents work from and cite.

  • Hybrid local model with a private tunnel

    Light models run where your people work. Heavy work goes through a private tunnel to infrastructure you can name.

  • Learning loop on open-weight models

    Open-weight models that improve from your organization’s own results. Learning continuously without forgetting is still an open research problem, and we treat it as research.

Bio-inspired systems and programs.

  • Vagus

    The vertical’s memory. Every claim keeps its date, source and authority, so a team can ask what was true on the day it mattered.

  • Manna

    Continuing work with a declared purpose. It keeps going between conversations and holds anything above its limits for a person.

  • Synapse

    The boundary for model work. It decides what an outside model may see and remembers what each provider has already received.

  • The Vault

    The method and component atlas the programs come from. Each method carries what it assumes and what proved it.

Tier 2 · Programs

Programs you install, review and own.

A program is working practice made executable. The system proposes it from your own work, a person approves it, and every version and run is recorded. Programs are installed one at a time or as a cluster for your industry.

Skill
A capability an agent can call, written for your domain: draft a tender response, reconcile a ledger.
Hook
A check before or after an action: a policy test, a disclosure check, a sign-off.
Workflow
A case carried across people and agents, with its states and who decides at each one.
Loop
Work that keeps watch: a weekly review, a deadline sweep, a drift check on a forecast.
Script
A deterministic step the agents rely on: an import, an export, a calculation with its units.
  1. 01

    Proposed

    The system proposes a program from the work it has read. Nothing installs itself.

  2. 02

    Reviewed

    A named person approves it, edits it or declines it.

  3. 03

    Installed

    The approved version enters your stack and stays under your ownership.

  4. 04

    Recorded

    Each version and each run lands in the evidence record.

Clusters by industry

  • Project operations

    The first organization pack: seven programs that carry project work from brief to delivery.

  • Legal practice

    Matter intake, conflict checks, clause review against your jurisdiction’s rules, deadline watch.

  • Engineering firm

    Specification review, change control, supplier qualification, drawing-register checks.

  • Logistics operator

    Exception handling, carrier choice with reasons, customs documents, delay forecasts with Extrasensory.

03 / 03 · Tier 3 · Institutional

Decisions you can replay.

Built and delivered by Naevum Labs.

At institutional scale, the questions that matter are who approved each action and whether the same inputs would get the same answer again. Intelligence proposes. A deterministic kernel admits, refuses or holds each action, and records which.

  • Manna as the deterministic decision kernel

    Manna proposes the next action; a separate deterministic kernel decides whether it goes ahead, waits for a person or is refused. Manna itself runs today in a controlled environment.

  • L5 orchestration

    Institution-wide orchestration at the top of the agent autonomy scale proposed by Feng and colleagues, where the agent works without user involvement.

Rented suites, or a stack you own.

Connect first. Retire what you no longer need.
Rented suitesThe Naevum stack
Where data livesIn each provider’s data centers, under its terms. Three US providers hold most of Europe’s cloud market.7Starts from your own folders. Runs on hardware you choose: your own, a managed bridge on rented EU servers, or Metahive sites.
Who can change the workflowThe vendor sets the product and its roadmap; your team configures inside it.The stack proposes programs, skills and workflows, and your people approve each one before it is installed.
Record of AI actionsActivity logs in each vendor’s format, kept for each vendor’s period.Every consequential action in one hash-chained, Merkle-anchored, timestamped record you can verify.
AI Act evidenceEach provider documents its own part; the deploying organization assembles the rest.Built for the AI Act’s obligations: roles allocated, mechanisms built, evidence kept. Records structured for Article 12.
Per-user license trendMicrosoft raised most Microsoft 365 commercial list prices on 1 July 2026, paid at each customer’s next renewal.3Connects to Microsoft 365, Google Workspace and Salesforce first, so a seat is retired only once the work runs without it.
Exit costEU cloud providers may charge switching costs until 12 January 2027; workflows built inside a suite do not move with the data.5Every fact in the knowledge base points back to a file you already hold, and the record verifies with standard Merkle proofs.

Which dependency would you retire first?

Plan your migration