Meshly Build / Sovereign
Your board, your agents and your models all stay in the building.
A control plane can be self-hosted and still leak, because the agents it governs have to talk to a model somewhere. Meshly closes that gap.
Meshly ships the other two pieces as well: a coding agent that runs against your own models, and a box to run them on. Meshly Build, Meshly CLI, Meshly Data Stack and the Meshly AI appliance can all run inside one airgap, and after installation nothing in the path reaches the public internet.
Meshly hosts Meshly Build by default. A sovereign deployment uses the self-hosted option, which runs on your own hardware and your own database.
Section A-A, the whole path
Every hop stays inside your network.
Once installation is done, no hop in this path reaches the public internet.
| Part | Runs on | Does |
|---|---|---|
| Meshly Build | Your server or container | Board, gates, audit, MCP |
| Meshly CLI | A developer machine | Agent loop, permission gate, sandbox |
| Meshly AI appliance | Your rack | Coder and reasoning models behind one gateway |
In an airgap, Meshly CLI connects to your local Meshly Build instance and to a local model, which can be the Meshly AI appliance, your own vLLM server or any OpenAI-compatible endpoint you run. Meshly Data Stack can run in the same network.
The gap this closes
Self-hosting the board was only ever part of the problem.
The execution edge of Meshly Build needs an agent, and an agent needs a model. If the only model available is a hosted frontier API, the airgap stops at the server boundary and the agents sit outside it.
That is why this used to be handled as a concession: the platform works offline, but your agents still need an exception. That answer is no longer the honest one, because we can now supply the models too.
Server, network unplugged
| External call | Behaviour offline |
|---|---|
| Licence heartbeat | It fails silently. |
| Usage telemetry | It fails silently. |
| Portal calls | They fail silently. |
| Slack, OIDC | These are optional and off by default. |
No code path denies you functionality. When the server runs in a degraded state, a banner says so, and nothing gets switched off.
Runs offline indefinitely. The server runs offline with no time limit. Licence heartbeat, telemetry and portal calls all fail silently and degrade cleanly.
Verified in code. We verified airgap operation in the code itself, so you do not have to rely on a claim in a datasheet.
Every other dependency is optional. Every other external dependency is optional, is off by default, or degrades without denying you any functionality.
Two routes
You can take the exception, or you can take neither.
The agents need model access, and we support two answers to that.
- Route 1
- Proxied exceptionAgents reach an approved gateway, which makes this the simplest route. Many environments described as airgapped permit exactly this, and if yours does, we will tell you so.
- Route 2
- No egress at allThe appliance serves local coder and reasoning models behind one OpenAI-compatible gateway on your own hardware, so nothing leaves. Meshly CLI is the agent that talks to it.
An ordinary endpoint. The gateway is an ordinary OpenAI-compatible endpoint, so the agent does nothing unusual to reach it.
The same MCP client. Meshly CLI holds the same MCP connection to the board that Claude Code and Codex hold, so the board sees the same MCP client either way. Gates, audit and cost accounting stay exactly as they are.
Hardware matters. Reference hardware exists for the appliance, and configurations outside it need tuning. Ask us before you assume yours fits.
Who this is for
We built this for buyers whose data must stay on their own ground.
When data sovereignty is a hard constraint for you, this stops being a conversation about features. Regulated and sovereignty-bound organisations used to be the hardest fit for this category. They are now the best fit, because the constraint that ruled everything else out is the one this was built around.
Meshly Oy is a European company, so for European buyers the vendor sits inside the same jurisdiction as the requirement.
Was a disqualifier, now a qualifier
If code, prompts and model traffic may not leave your building, that used to rule this category out for you. Today it is the reason to look at it.
Nothing leaves your network. Code, prompts and model traffic all stay inside your network, and no contract language has to do that work.
The record already exists. When governance obligations ask for auditable records of automated decisions, the record the board already keeps answers them.
Your own infrastructure. You can deploy on a single server or in containers, on your own hardware and your own database.
Stated honestly
This half of the stack is much younger.
We want you to read this here before you discover it in a pilot.
On 2026-11-18, Meshly Build was at v3.13.13 with 232 releases behind it, and Meshly CLI was at v0.8.1. The appliance was at 0.4.1 when we last read it, on 2026-08-20. The capability is real and in use, and we verified it in both source trees. It still carries far less production mileage than the control plane it serves. A customer who chooses the no-egress route is choosing the less-proven half, and should know that before signing anything.
| Product | Version | Status |
|---|---|---|
| Meshly Build | v3.13.13 | Mature, in production |
| Meshly CLI | v0.8.1 | Shipped, far newer |
| Meshly AI appliance | 0.4.1 | In use, built per customer |
We measured these versions on 2026-11-18, and the appliance on 2026-08-20.
Tool calls vary by model. Tool-call reliability with local models varies from model to model. Some work natively, and others need a parser fallback.
In use today. Others in this category describe this part as a roadmap. That difference is only worth something if we are equally plain about how new ours is.
Commercial terms are separate. Commercial terms for the appliance, including hardware, warranty and support, belong to a separate conversation, and we will not improvise them.
Contact
Tell us what your network allows.
The useful conversation starts with your actual constraint: what may leave, what may not, and who has to sign off. We will tell you which of the two routes fits, and if the simpler one is enough for you, we will say so.
One level deeper
You can check the boundary yourself.
The Trust Center covers deployment boundaries, security practices, and what is confirmed and what is still in progress.
Open the Trust Center