What exactly do you pre-install and configure?

The model server (Ollama or llama.cpp, Spark-tuned), OpenClaw wired to your channels and tools, the chat app with accounts for your team, and the tool-calling chat templates fixed for local models — the part the OpenClaw docs themselves flag as the common failure. Everything runs as supervised services with restart behaviour and logs.

Why do agents change the economics of owning the box?

Chat is bursty; agents are continuous. A looping agent on metered API pricing is a running tap — the same loop on owned hardware costs electricity. This is why the machine is sized and sold around token burn, not seats.

Can Claude Code really run against a local model?

Yes — two environment variables point it at any Anthropic-compatible server (Ollama 0.14+, LM Studio 0.4.1+). Caveats we state up front: no prompt caching, no batches, and tool-choice enforcement is looser than against Anthropic's own API.

Why isn't Cursor on the wall?

Cursor's own docs state all requests route through their servers for prompt building even with your own key and endpoint — and tab completion never leaves their models. On a machine sold on 'nothing leaves the building', that logo would be false advertising.

Pre-installed

Agents that work while you don't

We ship the box with the agent stack wired: OpenClaw connected to your channels, the runtime supervised, the tool-calling templates fixed for local models. Below is the ecosystem it plugs into — every 'fully local' badge verified against the tool's own docs.

Agent frameworks

What runs the loops. OpenClaw is the most-starred agent framework in local AI — and NVIDIA's own NemoClaw stack exists precisely to run OpenClaw and Hermes on this hardware.

OpenClaw

OpenClaw Foundation

Fully local

A persistent AI assistant across WhatsApp, Telegram and desktop that performs real tasks — the most-starred agent framework in local AI.

Local models may emit tool calls as raw text — fix the chat template; we pre-configure this.

MIT · 385k★

Hermes Agent

Nous Research

Partial

A self-improving agent that writes its own skills and keeps memory across sessions.

The agent loop runs fully local; the 60+ built-in tools route through Nous's paid cloud portal.

MIT · 225k★

n8n

Fully local

Visual automations across 400+ integrations that call your model as a node.

Fair-code · 199k★

Langflow

Fully local

Drag-and-drop agent builder that exports straight to an API endpoint.

MIT · 153k★

Dify

LangGenius

Fully local

A full LLM app platform — workflows, RAG, agents — self-hosted.

Modified Apache-2.0 · 151k★

CrewAI

Fully local

Role-playing agent crews — researcher + writer + reviewer as a team.

MIT · 57k★

Flowise

FlowiseAI

Fully local

Visual LangChain builder with a native local-model node.

Mixed OSS · 55k★

LangGraph

LangChain

Fully local

Stateful, resumable agent graphs with checkpointing — the serious way to build reliable agents.

MIT · 39k★

The sidekick pattern, which we set up by default: the flagship model does the thinking, while Liquid AI’s LFM2.5-2.6B — a ~3GB agentic model trained with agentic RL, vendor-reported ahead of models four times its size on tool use — stays resident and handles the cheap tool-call turns. The big model reasons, the small one runs the errands, and both live entirely on your box.

Coding agents

Point them at the box, keep your codebase in the building.

Zed

Zed Industries

Fully local

A fast native editor whose agent and inline assistant run off your own model — no account needed.

Apache-2.0 + GPL parts · 88k★

OpenHands

Fully local

An autonomous dev that opens a sandbox, writes code, runs it, and fixes its own errors.

Set OLLAMA_CONTEXT_LENGTH ≥ 22000 — the default is too small.

MIT · 83k★

Cline

Fully local

A VS Code agent that plans multi-file changes and asks approval at each step.

Declare context window and tool support manually for local models.

Apache-2.0 · 66k★

Goose

Block

Fully local

CLI + desktop agent that installs, executes and tests — extensible via MCP.

Apache-2.0 · 52k★

Aider

Aider AI

Fully local

A terminal pair programmer that git-commits every change so you can always roll back.

Apache-2.0 · 48k★

Continue

Fully local

Autocomplete, chat and edit in VS Code/JetBrains — a different model per role.

Apache-2.0 · 35k★

Kilo Code

Kilo

Fully local

The agentic VS Code extension merging the Cline/Roo lineage — dedicated local-model docs.

MIT · 27k★

Claude Code

Anthropic

Fully local

Anthropic's own CLI agent, pointed at your box with two environment variables.

Needs an Anthropic-compatible server (Ollama ≥ 0.14 / LM Studio ≥ 0.4.1); no prompt caching or batches locally.

Proprietary

Still needs the cloud — read before assuming

The claims above are believable because this list exists. These are the popular tools whose local support is partial or absent, in their own documentation's words.

CursorTheir own docs: “all requests are routed through Cursor's servers for final prompt building” — even with your key, your endpoint. Tab completion never leaves their models, and localhost endpoints don't work because their backend must reach yours.
Windsurf / DevinCloud SaaS only — no custom endpoint, no BYOK at all.
GitHub Copilot completionsBYOK chat works locally with no account; code completions and embeddings still require GitHub's cloud.
Hermes built-in toolsThe agent runs locally; its 60+ bundled tools (search, browse, vision) bill through Nous's portal. We wire local equivalents instead.

Questions, answered straight

What exactly do you pre-install and configure?: The model server (Ollama or llama.cpp, Spark-tuned), OpenClaw wired to your channels and tools, the chat app with accounts for your team, and the tool-calling chat templates fixed for local models — the part the OpenClaw docs themselves flag as the common failure. Everything runs as supervised services with restart behaviour and logs.
Why do agents change the economics of owning the box?: Chat is bursty; agents are continuous. A looping agent on metered API pricing is a running tap — the same loop on owned hardware costs electricity. This is why the machine is sized and sold around token burn, not seats.
Can Claude Code really run against a local model?: Yes — two environment variables point it at any Anthropic-compatible server (Ollama 0.14+, LM Studio 0.4.1+). Caveats we state up front: no prompt caching, no batches, and tool-choice enforcement is looser than against Anthropic's own API.
Why isn't Cursor on the wall?: Cursor's own docs state all requests route through their servers for prompt building even with your own key and endpoint — and tab completion never leaves their models. On a machine sold on 'nothing leaves the building', that logo would be false advertising.

Get the stack installed