VCN #41 · Bare Metal · 2026-07-15 · Frontier Tower F10

● tokens billed: $0.00

BARE METAL.

Run a coding agent against a local open-weights model. No API bill, fully private, your model on your metal.

WED JUL 15 Frontier Tower F10 Doors 7:00 Walkthrough 7:30 Season opener · #41

$10 early · $20 door · FT members free (FTMEMBER)

● doors 7:00 · walkthrough 7:30

WHILE YOU SETTLE IN.

wi-fi

network frontiertower

password frontiertower995

start the download now (~4.7GB)

brew install ollama
ollama pull qwen2.5-coder:7b   # start it NOW, ~4.7GB
# optional: score your machine
brew install AlexsJones/llmfit/llmfit && llmfit

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5-coder:7b   # start it NOW, ~4.7GB
# optional: score your machine
brew install AlexsJones/llmfit/llmfit && llmfit

# download + run the installer from ollama.com
ollama pull qwen2.5-coder:7b   # start it NOW, ~4.7GB
# optional: score your machine (PowerShell)
scoop install llmfit; llmfit

# enable Linux first: Settings > About ChromeOS > Developers
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5-coder:3b   # CPU-only, start small
uv tool install -U llmfit && llmfit

once it serves: run the wire one-liner from LAB 2.

no GPU? dev.nebius.com/builders for $50 free Nebius credits.

have a coding agent already?

Point it at https://vcn-41-bare-metal.vercel.app/setup.txt and tell it: set me up for tonight.

join Telegram Luma

● the problem · why this matters now

EVERY AGENT YOU RUN IS RENTING THE CORD.

Every coding agent you run today phones home to someone else's GPU and meters you per token. That is two costs, not one. [EV]

cost / 01

The bill

Per token, forever. Every run has a price, and the price never goes to zero while the cord is plugged in.

cost / 02

The leak

Your code, your prompts, your repo leave the building. Privacy is not a setting when the model lives on someone else's metal.

Tonight we cut the cord. The agent points at a model on your own metal. Zero per-run cost, nothing leaves the room.

● why now

OPEN-WEIGHTS CODERS GOT GOOD.

The open-weights coder tier is now genuinely usable for real tasks. These all live on Hugging Face. [HF]

● Qwen3-Coder ● GLM ● DeepSeek-Coder ● Llama

Serving is a solved commodity. Ollama, LM Studio, and vLLM all expose the same /v1 shape. [OLLAMA-OAI]

So the only question left is wiring. That is the whole night.

● concept · anatomy of the cord

ONE CONTRACT. SWAP WHAT ANSWERS IT.

Three boxes: coding agent connects to an OpenAI-compatible /v1 endpoint, which connects to an open-weights model on your GPU. The contract is POST /v1/chat/completions. Swap what answers it and the agent never knows.

The contract between agent and model is just POST /v1/chat/completions. Local vs hosted is nothing but which URL the agent's base points at. [OLLAMA-OAI]

● concept · pick one of each

THREE SERVERS. THREE AGENTS.

servers (pick one)

Ollama :11434/v1 (easiest)

LM Studio :1234/v1 (GUI, Start Server)

vLLM :8000/v1 (fastest, needs real VRAM)

agents (pick one)

Aider speaks OpenAI-compatible natively

Cline (VS Code) OpenAI Compatible provider

Claude Code native to Ollama since Jan 2026, no proxy

Reference path tonight: Ollama + Aider (fewest moving parts). Cline for a GUI. Claude Code talks to Ollama natively now. [OLLAMA-OAI] [CLINE] [OLLAMA-CC]

● concept · size to the metal

DO NOT GUESS YOUR VRAM. SCORE IT.

score your machine with llmfit

brew install AlexsJones/llmfit/llmfit

llmfit          # interactive TUI, ranks models for your machine
llmfit fit      # same ranking, plain list
llmfit doctor   # hardware report (ram / cpu / gpu)

brew install AlexsJones/llmfit/llmfit

llmfit          # interactive TUI, ranks models for your machine
llmfit fit      # same ranking, plain list
llmfit doctor   # hardware report (ram / cpu / gpu)

scoop install llmfit

llmfit          # interactive TUI, ranks models for your machine
llmfit fit      # same ranking, plain list
llmfit doctor   # hardware report (ram / cpu / gpu)

uv tool install -U llmfit   # inside the Crostini Linux terminal

llmfit          # interactive TUI, ranks models for your machine
llmfit fit      # same ranking, plain list
llmfit doctor   # hardware report (ram / cpu / gpu)

qwen2.5-coder: a size for every laptop

model

on disk

qwen2.5-coder:0.5b

398MB

qwen2.5-coder:1.5b

986MB

qwen2.5-coder:3b

1.9GB

qwen2.5-coder:7b

4.7GB · reference

qwen2.5-coder:14b

9.0GB

qwen2.5-coder:32b

20GB

The 3b runs on nearly anything and still closes simple tasks. Step up to qwen3-coder:30b (MoE) or deepseek-coder / GLM when you have the VRAM. Clean up with ollama rm <model>. [OLLAMA-LIB]

Apple Silicon: memory is unified, so VRAM is just your RAM. Check it via the Apple menu, About This Mac, Memory. 16GB or more comfortably runs the 7B (it needs about 8GB).

Pull the model llmfit ranks best for your metal, get the loop working end to end, then size up. Under-powered laptop is fine, that is the Nebius fallback. [LLMFIT] [OLLAMA-LIB] [EV]

● the replay · watch it close a task

$0.00

tokens billed

● tests PASSED · local agent · nothing left the room

A local agent reads a real repo, edits files, runs the test, and it goes green. Same agent UX you already know. No network egress. No invoice. This is the thing you will have working by 9:30. [EV] [OLLAMA]

● LAB 1 / 5 · stand up the server

PULL A MODEL. SERVE /v1.

Ollama is the reference path. Install, pull a coder model sized to your VRAM, serve the OpenAI-compatible endpoint, smoke test it. [OLLAMA] [OLLAMA-OAI]

Pre-pulled at home? 90 seconds. If not, this is the download slide (30 people on the event Wi-Fi). Be patient, or jump to LAB 5 and pull the local model later.

brew install ollama
ollama pull qwen2.5-coder:7b     # ~8GB VRAM, safe default
ollama serve                     # OpenAI-compatible at http://localhost:11434/v1
# smoke test the endpoint speaks /v1:
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen2.5-coder:7b","messages":[{"role":"user","content":"hi"}]}'

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5-coder:7b     # ~8GB VRAM, safe default
ollama serve                     # OpenAI-compatible at http://localhost:11434/v1
# smoke test the endpoint speaks /v1:
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen2.5-coder:7b","messages":[{"role":"user","content":"hi"}]}'

# download + run the installer from ollama.com
ollama pull qwen2.5-coder:7b     # ~8GB VRAM, safe default
ollama serve                     # OpenAI-compatible at http://localhost:11434/v1
# smoke test (curl is a PowerShell alias, use curl.exe; or run in Git Bash):
curl.exe http://localhost:11434/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"qwen2.5-coder:7b","messages":[{"role":"user","content":"hi"}]}'

# enable Linux first: Settings > About ChromeOS > Developers > Linux
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5-coder:3b     # no GPU in Crostini: CPU-only + slow, start small
ollama serve                     # OpenAI-compatible at http://localhost:11434/v1
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen2.5-coder:3b","messages":[{"role":"user","content":"hi"}]}'
# for a real task, prefer the Nebius fallback (LAB 5).

Mac gotcha: if ollama serve says "address already in use," the Ollama app is already running the server in the background. It looks like an error. It means you are done. Skip straight to the smoke curl.

Alternates. LM Studio: Start Server in the GUI, get :1234/v1. vLLM: vllm serve <model> --port 8000, get :8000/v1. [LMSTUDIO] [VLLM]

● LAB 2 / 5 · wire the agent to localhost

POINT THE BASE URL AT /v1.

Fastest path: the wire helper doctors your setup, installs and wires aider, then launches it. Or do it by hand below. Any string works as the API key. [AIDER-OLLAMA] [CLINE] [OLLAMA-CC]

fastest path: one command

curl -fsSL https://vcn-41-bare-metal.vercel.app/wire.sh | sh

curl -fsSL https://vcn-41-bare-metal.vercel.app/wire.sh | sh

powershell -ExecutionPolicy Bypass -c "irm https://vcn-41-bare-metal.vercel.app/wire.ps1 | iex"

curl -fsSL https://vcn-41-bare-metal.vercel.app/wire.sh | sh   # inside the Crostini terminal

or wire it by hand

# --- Aider (native OpenAI-compatible) ---
export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama_chat/qwen2.5-coder:7b

# --- Cline (VS Code) ---  Provider: OpenAI Compatible
#   Base URL: http://localhost:11434/v1   API Key: local (any)
#   Model ID: qwen2.5-coder:7b

# --- Claude Code (native to Ollama, no proxy; use a fresh terminal) ---
ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://127.0.0.1:11434 \
  ANTHROPIC_API_KEY= claude --model qwen2.5-coder:7b
# one-command alt: ollama launch claude --model qwen2.5-coder:7b
# revert: the vars above are one-shot; any new terminal runs hosted Claude Code

# --- Aider (native OpenAI-compatible) ---
export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama_chat/qwen2.5-coder:7b

# --- Cline (VS Code) ---  Provider: OpenAI Compatible
#   Base URL: http://localhost:11434/v1   API Key: local (any)
#   Model ID: qwen2.5-coder:7b

# --- Claude Code (native to Ollama, no proxy; use a fresh terminal) ---
ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://127.0.0.1:11434 \
  ANTHROPIC_API_KEY= claude --model qwen2.5-coder:7b
# one-command alt: ollama launch claude --model qwen2.5-coder:7b
# revert: the vars above are one-shot; any new terminal runs hosted Claude Code

# --- Aider (PowerShell env var syntax) ---
$env:OLLAMA_API_BASE = "http://127.0.0.1:11434"
aider --model ollama_chat/qwen2.5-coder:7b
# (or run in Git Bash and use the export form)

# --- Cline (VS Code) ---  Provider: OpenAI Compatible
#   Base URL: http://localhost:11434/v1   API Key: local (any)
#   Model ID: qwen2.5-coder:7b

# --- Claude Code (native to Ollama, no proxy; use a fresh terminal) ---
$env:ANTHROPIC_AUTH_TOKEN = 'ollama'
$env:ANTHROPIC_BASE_URL = "http://127.0.0.1:11434"
$env:ANTHROPIC_API_KEY = ''
claude --model qwen2.5-coder:7b
# one-command alt: ollama launch claude --model qwen2.5-coder:7b

# --- revert to hosted Claude Code: close this terminal, or ---
Remove-Item Env:ANTHROPIC_*

# --- run inside the Crostini Linux terminal ---
# --- Aider (native OpenAI-compatible) ---
export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama_chat/qwen2.5-coder:7b

# --- Cline (VS Code, Linux build) ---  Provider: OpenAI Compatible
#   Base URL: http://localhost:11434/v1   API Key: local (any)
#   Model ID: qwen2.5-coder:7b

# --- Claude Code (native to Ollama, no proxy; use a fresh terminal) ---
ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://127.0.0.1:11434 \
  ANTHROPIC_API_KEY= claude --model qwen2.5-coder:7b
# one-command alt: ollama launch claude --model qwen2.5-coder:7b
# revert: the vars above are one-shot; any new terminal runs hosted Claude Code

Trust gate: confirm the agent is actually hitting localhost (watch the ollama server logs) before you believe it is local.

● LAB 3 / 5 · drive a real task end to end

HAND IT THE TASK YOU BROUGHT.

Give the local agent your real task exactly as you would a hosted one. No task in hand? Add a failing test and make the agent turn it green. [EV]

cd ~/your-real-repo
aider --model ollama_chat/qwen2.5-coder:7b

# in the aider prompt:
#   /add src/thing.py tests/test_thing.py
#   make tests/test_thing.py pass; run the tests and fix until green
#   /run pytest -q

cd ~/your-real-repo
aider --model ollama_chat/qwen2.5-coder:7b

# in the aider prompt:
#   /add src/thing.py tests/test_thing.py
#   make tests/test_thing.py pass; run the tests and fix until green
#   /run pytest -q

# PowerShell or Git Bash, same aider commands:
cd ~/your-real-repo
aider --model ollama_chat/qwen2.5-coder:7b

# in the aider prompt:
#   /add src/thing.py tests/test_thing.py
#   make tests/test_thing.py pass; run the tests and fix until green
#   /run pytest -q

cd ~/your-real-repo    # inside the Crostini Linux terminal
aider --model ollama_chat/qwen2.5-coder:7b

# in the aider prompt:
#   /add src/thing.py tests/test_thing.py
#   make tests/test_thing.py pass; run the tests and fix until green
#   /run pytest -q

Watch it: read files, propose a diff, you approve, it runs the test, green. Note the token meter: 0 billed. [AIDER-OLLAMA] [EV]

● LAB 4 / 5 · local vs hosted, side by side

RUN THE SAME TASK BOTH WAYS.

You have z.ai + Claude Code in hand as the hosted baseline (GLM, provisioned for the session). Run the identical task against it. [EV]

local holds

Simple edits, refactors, boilerplate, anything private. Fast, free, on your metal.

local bends

Long context, multi-file reasoning, tricky bugs. Where the hosted model still earns its bill.

score honestly

# same task, both runners. compare:
#   did it finish?   how many turns?   quality of the diff?
#   latency?   dollars billed?

The takeaway is not "local wins." It is "now you know which one to reach for, per task." [EV]

● LAB 5 / 5 · fallback when the laptop can't host

SAME WIRING. SWAP THE URL.

Laptop too small for the model you want? Keep the exact same agent and /v1 wiring. Only the base URL and key change, to Nebius Token Factory (still open-weights, runs on your own free Builder Program credits). [NEBIUS-BUILDER]

# OpenAI-compatible. Only the URL + key change.
export OPENAI_API_BASE=https://api.tokenfactory.nebius.com/v1/
export OPENAI_API_KEY=<your-token-factory-key>   # from your Builder account
aider --model openai/<nebius-coder-model-id>
# Cline: LAB 2 screen, swap Base URL to that /v1/ endpoint.

# OpenAI-compatible. Only the URL + key change.
export OPENAI_API_BASE=https://api.tokenfactory.nebius.com/v1/
export OPENAI_API_KEY=<your-token-factory-key>   # from your Builder account
aider --model openai/<nebius-coder-model-id>
# Cline: LAB 2 screen, swap Base URL to that /v1/ endpoint.

# OpenAI-compatible. Only the URL + key change (PowerShell).
$env:OPENAI_API_BASE = "https://api.tokenfactory.nebius.com/v1/"
$env:OPENAI_API_KEY = "<your-token-factory-key>"  # from your Builder account
aider --model openai/<nebius-coder-model-id>
# Cline: LAB 2 screen, swap Base URL to that /v1/ endpoint.

# On Chromebook this is the realistic path (no local GPU).
export OPENAI_API_BASE=https://api.tokenfactory.nebius.com/v1/
export OPENAI_API_KEY=<your-token-factory-key>   # from your Builder account
aider --model openai/<nebius-coder-model-id>
# Cline: LAB 2 screen, swap Base URL to that /v1/ endpoint.

No key yet? dev.nebius.com/builders: $50 Token Factory credit, self-serve, about 5 minutes. Then browse coder models at studio.nebius.com and paste the exact id. [NEBIUS-BUILDER] [NEBIUS-MODELS]

● gotchas · trust gates

THE ONE THAT BITES FIRST.

gotcha 01 · the make-or-break setting

Context window: bump it, or the agent goes uselessly dumb.

Local models default to a tiny context (Ollama loads at 4k). The agent silently truncates your repo and looks broken when it is really just half-blind. Fix it once by baking a 32k variant of the model — shares weights, no re-download, no server restart. This one setting decides whether local feels good or feels dumb. [AIDER-OLLAMA]

printf 'FROM qwen2.5-coder:7b\nPARAMETER num_ctx 32768\n' > Modelfile
ollama create qwen2.5-coder:7b-32k -f Modelfile
aider --model ollama_chat/qwen2.5-coder:7b-32k   # Cline + claude too
ollama ps   # proof: CONTEXT must read 32768, not 4096
# VRAM is the price: an 8B loads ~5.3GB at 4k, ~9.1GB at 32k

printf 'FROM qwen2.5-coder:7b\nPARAMETER num_ctx 32768\n' > Modelfile
ollama create qwen2.5-coder:7b-32k -f Modelfile
aider --model ollama_chat/qwen2.5-coder:7b-32k   # Cline + claude too
ollama ps   # proof: CONTEXT must read 32768, not 4096
# VRAM is the price: an 8B loads ~5.3GB at 4k, ~9.1GB at 32k

Set-Content Modelfile "FROM qwen2.5-coder:7b"
Add-Content Modelfile "PARAMETER num_ctx 32768"
ollama create qwen2.5-coder:7b-32k -f Modelfile
aider --model ollama_chat/qwen2.5-coder:7b-32k   # Cline + claude too
ollama ps   # proof: CONTEXT must read 32768, not 4096
# VRAM is the price: an 8B loads ~5.3GB at 4k, ~9.1GB at 32k

printf 'FROM qwen2.5-coder:7b\nPARAMETER num_ctx 32768\n' > Modelfile
ollama create qwen2.5-coder:7b-32k -f Modelfile
aider --model ollama_chat/qwen2.5-coder:7b-32k   # Cline + claude too
ollama ps   # proof: CONTEXT must read 32768, not 4096
# CPU-only box? stay at 8k-16k, the window costs RAM

is it actually local?

Verify via server logs before trusting privacy claims. A misconfigured proxy can still phone home. [EV]tail -f ~/.ollama/logs/server.log # win: %LOCALAPPDATA%\Ollama\server.log # every request the agent makes prints here

tool-calling

Smaller models drop or malform tool calls. Keep a human approving diffs. Do not auto-apply. [EV]aider --no-auto-commits # review with /diff, back out with /undo

throughput

Ollama is convenient, vLLM is faster under load. VRAM OOM means drop a quant or a model size. [VLLM]ollama pull qwen2.5-coder:3b # OOM at 7b? step down a size, keep the loop

● leave with

WHAT YOU NOW HAVE WORKING.

A local loop

A coding agent wired to a local OpenAI-compatible endpoint, closing a real task with zero tokens billed.

A working rig

A server (Ollama / LM Studio / vLLM) plus an open-weights coder model, pulled and serving on your metal.

A read + a fallback

A side-by-side on local vs hosted, and a Nebius fallback for when your metal runs out.

By 10pm: an agent completing a real task, driven entirely by a local open-weights model. No bill. Nothing left your metal. [EV]

● resources

WHAT YOUR TICKET INCLUDES.

hosted baseline

z.ai + Claude Code, provisioned for the session (GLM). The hosted side of the comparison.

labs compute

Nebius Builder Program, self-serve: dev.nebius.com/builders gives you $50 Token Factory + $50 AI Cloud + $25 Tavily credit, in about 5 minutes.

repo + site

vibecodingnights.com for the commands, the model list, and the next event.

Your ticket includes z.ai + Claude Code. Nebius credits are self-serve via the Builder Program. [EV] [NEBIUS-BUILDER]

● sponsors · on topic for local open-weights serving

POWERED BY.

Ollama

The easiest local OpenAI-compatible server. The LAB 1 reference path.

Nebius Token Factory

The Builder Program gives every attendee $50 Token Factory + $50 AI Cloud + $25 Tavily credit, self-serve at dev.nebius.com/builders.

Hugging Face

Home of the open-weights coder models we pulled: Qwen3-Coder, DeepSeek-Coder, GLM.

● community · off the leash, #41 to #50

BRING A LAPTOP. TEACH A NIGHT.

One real thing each session, and you leave with it working. No slideware, no spectating.

join vcn

Telegram · t.me/+EBFzKXmJAVk5ZGU0 vibecodingnights.com luma.com/vcn-41-bare-metal

Two formats a week: Nights (Wed 7pm, Floor 10) and Mornings (Sat 10am, Floor 9).

RSVP the next one. Teach a night, builder-led talks welcome. Sponsor or co-host? Reach Facilitator Rayyan Zahid.

Hosted by Vibe Coding Nights: Rayyan Zahid (Immersive Commons), Michalis Vasileiadis (Hacker Bob), Eric Mockler (AI Geneticist), Devinder Sodhi (Learning Layer Labs).