BareMetalRT — Bare Metal AI

GPU<br>Fleet<br>Mesh

BARE METALRT

). WIP light CSS is<br>preserved in app.css. -->

 Sealed<br>OFFLINE

Sign In<br>Sign Up

Account

Admin

Sign Out

What's new<br>Update now<br>&times;

&times;

Skip tour<br>Next &rarr;

&times;<br>Ask Ada

Ask Ada

Help & FAQ for AI newcomers

&times;

Report a bug

💬 Live AI answers from Ada are coming soon

1)history.back();else location.href='/'">&times;

Bare Metal AI

Your AI, on hardware you own.

The local AI agent runtime for Windows — open-weight models and tool-calling agents on the GPUs you own. Private, entirely yours.

Prompts run on your GPUs — they never leave your network

We never see, store, or train on your data

On-prem & air-gap ready · SSO & SCIM for the enterprise

Built for regulated, private AI.

Sign In

Welcome back — sign in to your GPUs.

or use email

Sign in with SSO

Continue with Google

Continue with Microsoft

Continue with GitHub

Full name

Username

Username or email

Password

Show

Use at least 8 characters

Authentication code

Sign In

No account? Sign up

Forgot password?

Try the live demo &rarr;

Admin<br>&times;

Maintenance Banners

Bug Reports<br>Refresh

Leads & Applications<br>Refresh

+ New session

Memory

Loading…

Agents

No runs yet

+ New agent run

Routines

No routines yet

+ New routine

Skills

No skills enabled

+ Browse skills

Integrations

▾Built-in

Connected

Connect

GmailConnect

CalendarConnect

Google DriveConnect

Microsoft 365Connect

+ Connect a Git repo

+ Add MCP server

+ Browse integrations

Workflows

No workflows yet

+ New workflow

Sessions

No sessions yet

Your chat history is AES-256 encrypted on your device — never stored on our servers.

Privacy &middot; Terms &middot; v0.21.25

&copy; 2026 Bare Metal AI, Inc.

&larr;

Browse

Connect tools the agent can use. Everything runs on your GPU host — tokens and data stay local.

+ Add MCP<br>Request integration

Integrations<br>Skills

Loading…

&larr;

Audit Log

A tamper-evident, hash-chained record of every tool call, approval, connector change, auth event and config change. Stored only on your GPU host .

Checking chain…

All events<br>Tool call<br>Tool approved<br>Tool denied<br>Connector connected<br>Connector removed<br>Auth login<br>Auth failed<br>Model loaded<br>Config changed

Export CSV<br>JSON

#Time (UTC)SourceEventTargetDetailResult<br>Loading…

&larr; Newer

Older &rarr;

&larr;

Memory

What the agent remembers about you across chats. Stored only on your GPU host — never sent to any server. Enabled items are added to the start of every chat. Since it lives on this machine, export a backup to keep it safe across reinstalls.

Connect apps<br>Open in Obsidian<br>Export<br>Import<br>+ New

Loading…

&larr;

Agents

Headless agent runs on your GPU host — give a goal, watch it work. Runs queue and execute one at a time on the engine. Read-only refuses side-effecting tools; autonomous is pre-authorized to use them. The decision is made at launch — there is no mid-run approval.

Read-only — safe, no side-effecting tools<br>Autonomous — can use all tools

Launch run

No runs yet.

&larr;

Routines

Scheduled agent runs on your GPU host — always-on, no cloud bill. Each routine launches an agent run on its schedule; the runs appear in the Agents fleet. Everything stays on this machine.

Read-only<br>Autonomous

Every N hours<br>Daily at…

Add routine

No routines yet.

&larr;

Workflows

Build a repeatable process as a funnel of steps — deterministic tool actions piped into AI reasoning nodes — that runs on a model you choose, entirely on your GPU host. Same inputs, same path: every run is recorded so you can audit and improve it.

📘 How workflows work — node types, placeholders & determinism

A workflow is a repeatable process you build as a vertical funnel : inputs at the top, ordered steps narrowing down, and a deliverable at the spout. The daemon walks the funnel in order and pipes each node's output into the next — control flow is enforced by the engine, not improvised by the model, so a workflow is repeatable and auditable (Power-Automate-style).

The four node types

Input A seed value — the object/data the process starts from (e.g. a query, an ID, a block of text).

Action A deterministic tool call. Pick a tool (a built-in or a connected integration) and bind its parameters. Runs exactly as specified — no model involved.

AI A reasoning step. Write an instruction and the chosen model produces text (summarize, classify, draft). The only place the model decides anything.

Output The deliverable — captures the previous node's result as the workflow's final output.

Piping outputs forward — placeholders

Every node has a short id (shown on its card: n1, n2, …). Reference an earlier node's output anywhere in a later node's instruction or action parameters with {{n1}}. Example: an AI node with Summarize {{n1}} and {{n2}} into a short report receives the live outputs of nodes 1 and 2.

Model & permissions

The Model field picks which model runs the AI nodes; the daemon auto-swaps to it before the...