Your own AI, on your own hardware.
A compact on-prem appliance that runs local models in three modes: reasoning, a coder brain, and an artist mode for image generation. Your data, prompts, and models never leave the box. No calls to a public AI service. Two ways in: Meshly CLI in the terminal, and Ilves, the desktop chat and Image Studio app.
Reasoning, coder, and an artist.
One box, three brains. Switch between a careful reasoning model, an agentic coder, and on-box image generation. It runs one heavy model at a time and re-routes on demand, so the whole appliance is always pointed at the job in front of it.
- Reasoning: depth-first thinking for triage, analysis, and knowledge work.
- Coder: an agentic brain that writes, edits, and runs your code.
- Artist: prompt-to-image generation that never touches the cloud.
In coder mode, it runs Meshly Build and Data Stack.
The coder brain doesn't just write code. It can operate Meshly Build and Meshly Data Stack for you, even fully airgapped. Move tasks, query data, run the agent loop: your own AI driving your own stack, with every call staying inside the perimeter.
- Operates Meshly Build: plan, delegate, move tasks, push results.
- Queries Meshly Data Stack through on-box tools, no external calls.
- Works fully airgapped. External connections are off by default on the box.
Meshly CLI and Ilves.
Talk to the box however suits the work. Meshly CLI gives engineers a terminal-native agent. Ilves is the desktop app: a chat client and Image Studio for macOS and Windows. Same appliance underneath, two front doors.
- Meshly CLI: terminal agent for coding and operating your stack.
- Ilves: desktop chat + Image Studio for macOS and Windows.
- Model-neutral: bring your own open or local models, no per-seat or per-token metering on the box.
The agent CLI that doesn't need the internet.
Meshly CLI is the terminal interface to the box. Point it at your appliance, or any OpenAI-compatible endpoint your team runs, and it gives engineers the muscle memory of a modern coding agent with the model staying in-house. The agent inside is called Ilves.
- Same tool surface as Claude Code: Read / Edit / Write / Bash / Glob / Grep, plan mode, MCP.
- Workflow-compatible, not protocol-compatible. No public AI API in the picture.
- Standalone, or connected to Meshly Build to dispatch tasks and stream results back.
No phone-home. No public-internet calls.
Inference never leaves the machine. No telemetry, no auto-update pings, no anonymous usage stats. On a shipped appliance, external connections are off by default: you opt in, per box, with explicit allow-lists. Run it on a closed network and sleep at night.
- Inference stays on the box. No cloud fallback, ever.
- No telemetry or phone-home; the only traffic is what you configure.
- Airgap-verified: models pre-cached, runs with the network unplugged.
For the teams that can't send data to cloud AI.
Cloud AI works great until your compliance officer reads the data processing addendum. Meshly AI brings the model to the data, not the other way around. The whole loop runs on a box you own.
Available in early access, by request.
Meshly AI isn't a public download. Tell us about your environment: what hardware you have, what models you run, what restrictions you operate under, and we'll work out whether it fits and get you a box with the right configuration.