Give your GPU a job.
Install a Windows, Linux, or Apple Silicon macOS agent on your compatible machine. Connect the host to your server and bring its supported model runtimes into view.
YOUR HARDWARE. YOUR MODELS. YOUR RULES.
Agentic coding.
Your way.
Plan it. Build it. Approve every step.
Turn your GPU-powered PC, laptop, or Apple Silicon Mac into an LLM node with VLM Code Server. Then let VLM Code plan the project, research what already exists, build it phase by phase, and stop at every gate for your approval.
Local-firstResearch-backed plansPhase gates you approve
THE VLM ECOSYSTEMArchitecture overview
Your desktopGPU NODE / 01
Your laptopGPU NODE / 02
THE MODEL LAYERVLM Code Server
VLM CodeYour agentic workspace
ONE CONNECTED WORKFLOW
From the GPU on your desk to the code you ship. Now with Plan mode research, phase gates and a full toolbelt.
↓ THE TWO-MINUTE TOUR
Everything VLM Code does now.
Everything VLM Code does now.
In two minutes.
From the GPU on your desk to an approved, working phase: your own model nodes, research-backed planning, the build itself, and the gates that keep you in charge.
01 / VLM CODE SERVERYour next AI node
Your next AI node
could be right here.
A desktop. A laptop. An Apple Silicon Mac.
Give it a new role in your AI workflow with the VLM Code Server Agent.
BRING YOUR OWN COMPUTE
Every node starts with your hardware.Windows · Linux · macOS on Apple Silicon
Your models. Your infrastructure.
See host health, GPU and memory telemetry, and model status in one dashboard. Start, stop, and inspect supported model containers and native MLX services through the agent.
Connect it. Start creating.
Create a scoped API key, connect the server’s OpenAI-compatible endpoint in VLM Code, and select your model. One connection brings your nodes into your coding workflow.
Model size, GPU or unified memory, operating system, and runtime support determine what each node can run.
NEW / APPLE SILICON + MLX
Your Mac. Now an AI node.
Run models natively on Apple Silicon with MLX, using Metal acceleration and your Mac’s unified memory. The macOS agent manages native MLX services—no Docker Desktop required.
Connect through the same VLM Code Server API and bring your Mac’s inference into VLM Code, alongside your Windows and Linux nodes.
One API. Your nodes.
Scoped keys control which agents each client can use, with a shared OpenAI-compatible API for model discovery and inference.
A model, or your Router.
Choose a specific model or use the Router’s ordered targets to select the first available model your key can access.
Know your infrastructure.
Follow host heartbeats, runtime state, resource readings, and command results from the management dashboard.
02 / VLM CODEMore than a conversation.
More than a conversation.
A workspace that acts.
Give your agent the context, tools, and model it needs.
Then follow the work from the first plan to the final change, with every action on screen.
Your model. Your workflow.PLAN → RESEARCH → BUILD → VERIFY → APPROVE
Grounded in your project.
Workspace-aware file tools, project search and a built-in file explorer keep your code at the centre. Open any file the agent touched, rendered or as source.
Tools that do the work.
Read and edit files, run builds and tests, drive a browser, and watch every process. The agent verifies its changes before it calls anything done.
Make the workflow yours.
Choose your model and your helper model, add skills, plugins and MCP tools, and decide which actions need your approval.
03 / PLAN MODEPlans like an architect.
Plans like an architect.
Researches like an engineer.
Describe the project. Plan mode sizes it, researches what already exists in the .NET ecosystem, and writes linked development manuals with a phase plan you approve step by step.
01 / SIZE
Know how big it is.
Every plan starts with an honest size: small, medium or large. Large projects are always split into phases, because local models do their best work in focused sessions.
02 / RESEARCH
Check before it chooses.
The agent looks up real NuGet packages (license, supported frameworks, deprecation), finds open-source projects on GitHub and reads official docs with web search and the crawler, then cites what it found.
03 / PHASES & GATES
Build in steps you approve.
Each phase ends in something you can run, with the checks the agent verifies and the checks you perform. At every gate: approve, request changes, or stop.
CANDIDATELICENSEDECISION
MarkdigBSD-2ADOPT
Markdig.WpfMITCONSIDER
Custom parser—REJECT
Risk check: WinUI 3 needs extra tooling here. Recommend WPF on .NET 10.
Sources: nuget.org · github.com · learn.microsoft.com
- 1Walking skeletonApproved · window opens, tests pass
- 2Open & renderIn progress · open any .md file
- 3Recent files & polishNext · last five files remembered
The manuals, the research table and phase-plan.json live in your workspace under Development-Plans. The Phases tab shows where every phase stands.
NEW IN VLM CODESmarter plans.
Smarter plans.
Smoother flow.
Plan mode that researches
Sizes the project, checks real NuGet packages and GitHub projects, and cites its sources.
Phase gates
Large builds are split into runnable phases. The agent stops at each gate for your approval.
File explorer & viewer
Browse everything the agent touched and open files in place, rendered or as source.
Suggested replies
Your likely next message waits in the composer. Press Tab to use it.
Smarter Improve prompt
Turns "fix that error" into a precise instruction using the conversation so far.
Durable, resumable runs
Long runs survive interruptions and pick up from their checkpoints.
04 / THE TOOLBELTA full toolbelt.
A full toolbelt.
Built in.
A capable agent needs more than a model. Every action is visible, approvable and resumable, with the tools real development needs.
VERIFY
Agent Browser
The agent tests your local web app in its own Chromium window: it clicks, types, takes screenshots and reports console and network errors.
OBSERVE
Process Tracker
Every command the agent starts, with live status, bounded output, exit codes, stop controls and exact restarts.
PREVIEW
Blaze previews
Built-in local previews for static sites and Blazor WebAssembly apps, supervised alongside your other processes.
VERSION
Git built in
Status, diffs, branches and commits from the same conversation that made the change.
OPERATE
SSH to your servers
Operate remote Linux and Windows hosts with credentials held by the host for the session, not pasted into chat.
MOBILE
Android toolkit
Build, deploy, inspect and test Android apps on emulators and devices through the SDK you already have.
EXTEND
MCP servers
Web search, a site crawler, document tools and any MCP server you add, with per-server control.
EXTEND
Plugins & skills
Install Agent Plugins that bundle skills and MCP tools, and switch each plugin on or off.
DELEGATE
Helper agents
Hand focused tasks to helper agents, even on a different endpoint such as Codex, while your local model leads.
DESIGN
Designer
Design and preview interfaces visually in a dedicated design workspace.
AUTOMATE
Mail agent & schedules
Let the agent triage mail and run recurring jobs on a schedule you set.
ANYWHERE
Phone remote
Follow runs, answer questions and approve actions from your phone over your local network, behind a password.
HARDWARE TO HELLO WORLDOne connected workflow.
One connected workflow.
Entirely your own.
Bring a GPU.
Set up the Server Agent on a compatible PC, laptop, or Apple Silicon Mac, and connect its OpenAI-compatible endpoint to VLM Code.
Describe the goal.
Plan mode sizes the work, researches packages and existing projects, and writes manuals with a phased plan.
Let the agent work.
It writes code, runs builds and tests, checks the result in a browser or emulator, and shows every step as it goes.
Open each gate.
Try the phase yourself. Approve it, ask for changes, or stop. Your next reply is already suggested; press Tab.
LOCAL-FIRST. CHOICE ALWAYS.Your code.
Your code.
Your compute.
Your call.
Own the environment behind your ideas.
Keep inference on your machine or connect to a node you control. VLM Code keeps sessions, settings and workspace state on your computer; prompts and relevant context go only to the model endpoint you choose.
- Approvals your way. Auto-approve everything, only commands, only workspace edits, or nothing at all.
- Gates you open. Phased builds stop for your validation before the next phase starts.
- Nothing hidden. Every tool call, process and changed file is visible, and runs resume from their checkpoints.
- From your phone, on your terms. The optional phone remote stays on your local network, behind HTTPS and your password.
VLM CODE + VLM CODE SERVERThe power is yours.
The power is yours.
Build something great.
Your hardware. Your models. A plan you approve, one phase at a time.
Commercial software · VLM Code for Windows