Kosa8 Labs stack
Model, sandbox, governance. One sovereign stack.
Kosa8 Labs makes three products that fit together. llamay runs the model on your hardware. kosa8 runs the agent in a microVM you can fork. MAQPNA decides what the agent may do and keeps one record of all of it.
Three products, one job: agents you can run on your own terms
MAQPNA
Governs. Every tool and model call goes through one gateway that checks identity, policy, data rules and budgets, asks a person when needed, and writes a tamper-evident record.
llamay
Thinks. A local inference engine for open models (GGUF), on CPU or GPU. Its context is an object: it can be forked, saved and restored, so a long conversation is not re-read on every turn.
kosa8
Contains. Every sandbox is its own microVM, and a running sandbox can be forked into live copies in about a second. It speaks the Docker CLI and API, so your images work as they are.
How it fits
One call, three layers.
- 01
The agent runs in kosa8
A microVM whose only network destination is the MAQPNA gateway.
- 02
MAQPNA decides
Identity, policy, DLP, budget and approval, for every tool and model call.
- 03
llamay answers
The model call goes to a local model, in the session's own context.
- 04
One record
llamay's and kosa8's own audit logs are anchored in MAQPNA's ledger, so one check covers all three.
What you get together
More than the sum of three tools.
A model context per session
Each agent session gets its own llamay context. Fork the session and the context forks with it.
Fork a live agent
Copy a running sandbox into governed child sessions to try alternatives side by side.
One chain of evidence
Model requests, sandbox events and decisions verify end to end.
Nothing leaves
No hosted model, no API key, no phone-home. Run it offline.
Choose
Use one, two or all three.
| You run | Good for | Where |
|---|---|---|
| MAQPNA | Governing agents that use the models and sandboxes you already have | Laptop and Kubernetes |
| MAQPNA + llamay | Agents on private models, with a context per session and a request audit | Laptop and Kubernetes (llamay ships in the MAQPNA Helm chart) |
| MAQPNA + kosa8 | Local agents in microVMs that reach only the gateway, with live fork | Laptop and edge machines |
| All three | Fully private agents: local model, microVM boundary, one verifiable record | Laptop and edge machines; in a cluster, MAQPNA's own sandbox tiers take kosa8's place |
macOS on Apple silicon is the tested path for kosa8 with MAQPNA. Linux with KVM uses the same kosa8 API. Details: compatibility.
Try it
Four commands to a private agent.
# install the three CLIs
curl -fsSL https://maqpna.com/install.sh | sh
curl -fsSL https://llamay.com/install.sh | sh
curl -fsSL https://kosa8.com/install.sh | sh
# a small model, the sandbox daemon, then MAQPNA with both
llamay pull qwen2.5:0.5b
kosa8 daemon start
maqpna dev up --llamay qwen2.5:0.5b --kosa8
maqpna dev run --kosa8-image python:3.12-alpine -- python -c "$(cat agent.py)"Do I need all three?
No. MAQPNA works with any OpenAI-compatible model and its own Kubernetes sandboxes. Add llamay or kosa8 when you want what they add.
Who makes them?
Kosa8 Labs makes all three.
Does llamay need a GPU?
No. Small models run on a laptop CPU. llamay also runs on CUDA GPUs.
Can I run the stack in Kubernetes?
MAQPNA and llamay, yes: the MAQPNA Helm chart can deploy llamay as the model backend. kosa8 is for laptops and edge machines; in a cluster, MAQPNA isolates sessions with gVisor, Kata microVMs or confidential VMs.
All three on one laptop in about fifteen minutes.
A local model, a microVM sandbox and a governed agent. Nothing leaves the machine.
curl -fsSL https://maqpna.com/install.sh | shirm https://maqpna.com/install.ps1 | iex