23 real agent packs

Your team already uses AI. You just can't see it.

68% of employees use personal AI accounts for work — 22% even when the company gives them a tool. That's your data, leaving without a record. MAMA brings the work into one governed harness: You decide what leaves. You can see everything that did.

Master Autonomous Management Agent

Join Private Waitlist See the board
23 real agent packs | Morning brief on a schedule you set | Private tier targets an 8 GB laptop — not a GPU rack | Slack live · Asana · HubSpot, Gmail & more rolling out
MAMA · The Board EXAMPLE
How a day on the board looks

A representative board: agents pick up tasks, collaborate, and ship.

To do3
Fix greet() format
MAMA Architect brief
Wire Slack brief enrollment
MAMA Architect brief
Stripe webhook mount order
MAMA Architect brief
In progress2
Seal PHI routing table
MAMA Architect brief running
Gateway cost pass
MAMA Builder
Review1
Attestation PDF export
MAMA Builder tests ✓
Done2
Loopback bind + bearer auth
MAMA Builder verified ✓
Refusal ledger
MAMA Builder verified ✓

●  live product · mama.oliwoods.ai

Dashboard Agents Goals Marketplace Content Team TNT Voice

Dashboard

The problem
Shadow AI isn't coming.
It's already on payroll.

68% of employees use personal AI accounts for work. 22% keep doing it even after the company hands them a tool. Every one of those prompts is company data leaving the building — with no record that it left.

The leak is already open

Contracts pasted into personal ChatGPT. Client lists summarized on a free account. It happens every day, in good faith, and nothing is logged.

Banning it doesn't work

One in five employees keeps using personal AI even when handed a company tool. Prohibition doesn't stop the work — it just pushes it further out of sight.

You can't govern what you can't see

No record of what left, where it went, or what it cost. The answer isn't another ban — it's a place where the work is better, and everything is on the record.

How it works

Three steps. Then the harness runs itself.

Connect your tools. Install your agents. MAMA manages the rest — autonomously, every day.

1

Connect

Install MAMA in Slack — that is the live surface today. Asana connects with a token you paste. HubSpot, Gmail, and Calendar are rolling out next. No engineering required.

Slack live · Asana token · More rolling out

2

Install Agents

23 real agent packs across sales, finance, research, and legal. Or upload your own custom agent.

23 real packs · Sales, Finance, Research, Legal

3

Let MAMA Run

Morning briefs on your schedule. Tasks routed to agents. Costs tracked automatically. You review, approve, and stay in control.

Scheduled brief · Auto-routing · Budget caps

Inside the harness

What's inside the harness.

Morning Brief

A scheduled Slack post with pipeline, deadlines, open PRs, and agent spend — what matters today, when you want it.

Agent Marketplace

23 real packs. Sales, finance, legal, research. Install in one click. Each pack ships with pre-tuned prompts and tool access.

You Decide What Leaves

Claude, GPT, Gemini, Groq — bring your own models, routed per task under rules you set. Every call is logged.

Cost Control

Daily budget caps. Per-agent spend tracking. Full visibility, always. And work that stays on your own machine is priced at what it actually costs to run — nothing.

Task Router

Describe the task. MAMA routes it to the right agent, tracks completion, and surfaces blockers in Slack before they become problems.

Slack-first today

Lives where your team already works. Web dashboard and more surfaces ship as you grow. One memory, one control plane.

iOS Widget

She keeps working while your phone is in your pocket.

Morning brief on your schedule. Priorities on your home screen. Prompt her from the lock screen — she routes, tracks, and reports back.

MAMA EXAMPLE
Good morning, Matt.
5 need attention · 47 done overnight
P0 CoFounder WebSocket scaling
P0 Stay22 API integration
P1 GRAIL alert notifications
P1 TNT PPG 22 agents deployed
Hey MAMA…
MAMA Private — in development · early access

You've been told
you can't use AI.

Law firms. Advisors. Clinics. Agencies under NDA. The rule is simple: the data cannot leave the building. Most "private AI" answers that rule with a bigger machine. We answered it with a smaller model — MAMA Private is built for the laptop your team already has.

01 — Residence

The machine you already own

Not a workstation. The router budgets real memory — model weights plus the KV cache at the context we actually serve — and picks the largest model that genuinely fits an 8 GB laptop.

02 — Consent

Your say-so

Nothing leaves your perimeter without explicit permission — and permission is granted per source, with no wildcards and no inheritance. Not a prompt, not a document, not a telemetry ping.

03 — Record

Your receipts

Every departure is logged: what left, where it went, and why. And when you purge, the database is vacuumed — deleted text does not survive in free pages.

The compute ladder — budgeted, not guessed

A declared 128k context window with no memory budget behind it is how small machines die. So every tier carries measured weights and a real KV-cache cost, and the selector takes the largest one that fits.

8 GB
MacBook Air · 4.5 GB usable
qwen3:4b
Endpoint
16 GB
Standard laptop · 12.5 GB usable
phi3:14b
Workstation
32 GB
Desk machine · 28.5 GB usable
qwen2.5-coder:32b
Brain
64 GB
Studio box · 60.5 GB usable
deepseek-r1:70b
Brain
Model Needs 8 GB 16 GB 32 GB 64 GB
llama3.2:1b 1.2 GB ✓ ✓ ✓ ✓
qwen3:1.7b 1.6 GB ✓ ✓ ✓ ✓
llama3.2:3b 3.2 GB ✓ ✓ ✓ ✓
qwen3:4b 3.9 GB ✓ ✓ ✓ ✓
mistral:7b 5.7 GB — ✓ ✓ ✓
phi3:14b 9.5 GB — ✓ ✓ ✓
qwen2.5-coder:32b 20.3 GB — — ✓ ✓
deepseek-r1:70b 43.5 GB — — — ✓

EXAMPLE Every figure is generated by the same function the router uses — measured Q4 weights, plus the KV cache at 8k context, plus runtime overhead, against total RAM minus a 3.5 GB macOS reserve. They are computed budgets, not a live reading from your machine. 7B is where most private AI starts, and it misses an 8 GB laptop by 27%.

Refusal 01

No model fits? It stops.

The selector returns nothing rather than quietly reaching for a cloud model to cover a small machine. The caller has to refuse.

Refusal 02

Can't honour the schema? It says so.

A provider with no structured-output support is refused, never silently downgraded — because "the model returned prose" and "we forgot to ask" look identical at the call site.

Refusal 03

No local embedder? It won't improvise.

Indexing your own documents refuses rather than falling back to a cloud embedder. Sensitive text does not leave to make a feature work.

Anyone can claim privacy. Fewer can show you where the software stops. MAMA Private is still in development and is not for sale, and no figure above is a live reading from a customer machine yet. We would rather promise you less and prove it, than promise everything and hope you never check.

Join the MAMA Private Waitlist

Early access — in order, no queue-jumping. No spam, ever; only word about MAMA Private.

You're in line. We'll write when early access opens.