local only $0 local no code leaves the machine

Just Fucking Use Open Code.

Stop Gift-Wrapping Your Company's IP for Strangers. Just Fucking Use Open Code.

OpenCode is MIT-licensed and runs in your terminal. It points at any model, keeps your diff on your disk, and answers to nobody's sales team. Pair it with a local model and your autocomplete stops moonlighting as someone else's training data.

opencode · local session

❯ opencode run --model qwen2.5-coder:1.5b

connected: localhost:11434 ● no api key, no account, no invoice

❯ refactor the auth module, touch nothing outside it

+ 3 files edited locally, 0 bytes sent anywhere

- skipped: uploading your repo to a stranger's gpu

❯ /init

AGENTS.md written, the agent now knows your repo better than your tech lead

localhost:11434 ● connected qwen2.5-coder:1.5b

Own Your Autocomplete

Renting Intelligence is for Cowards. Own Your Autocomplete. The weights live on your disk. Every keystroke resolves at home, where no product manager is A/B testing your semicolons.

Local Models Don't Leak

You are One Data Breach Away from Getting Fired. Local Models Don't Leak. No vendor to subpoena, no status page to refresh, no breach notification with your repo in it.

Run It Locally

Your API Bill is a Joke. Run It Locally. One download, about six gigabytes, then quiet. No seats, no tokens. Nobody invoices you for typing.

❯ Your API bill is a joke

Item Local The subscription
Price $0, forever $20/seat/mo + tokens, forever
Code path localhost:11434 a data center with your name on a slide
Works offline yes no, and the outage always lands mid-demo
Lock-in none, MIT-licensed your muscle memory, their retention metric

❯ What the fuck is OpenCode

An MIT-licensed agent that lives in your terminal, drives 75+ models including the ones on your own disk, and asks for nothing but a shell. Install it like software used to be installed:

$ curl -fsSL https://opencode.ai/install | bash

Any model, even yours

75+ providers, free ones in the box, local ones over there. A long menu with nobody upselling dessert.

Plan first, touch later

Build and Plan modes, Tab to switch. It proposes before it touches, and /undo exists for the times it touches anyway.

Every window you own

Terminal, desktop, IDE extension. Your editor gets a roommate that pays rent in completions.

Stores nothing

No code, no context, no telemetry to improve your experience. There is no experience. There is a terminal.

195,000 stars, 950 contributors, 16 million developers a month. Your peers have voted, and it was not close.

Want managed models anyway

OpenCode Go is the team's low-cost plan: tested open coding models, hosted, minus the hyperscaler baggage. Zen exists too, for the days local is not enough. Both optional. Local stays the default around here, and the page will judge you only a little.

Get OpenCode Go

❯ Four commands, fifteen minutes

$ brew install ollama
$ ollama pull qwen2.5-coder:1.5b
$ ollama pull qwen2.5-coder:7b  # chat and edits, only if you have the RAM
$ ollama pull nomic-embed-text  # codebase search

Then introduce OpenCode to Ollama in opencode.json. Yes, a config file. You have survived worse for a linter.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ollama (local)",
      "options": { "baseURL": "http://localhost:11434/v1" },
      "models": {
        "qwen2.5-coder:7b": { "name": "Qwen 2.5 Coder 7B (local)" }
      }
    }
  }
}

❯ Hardware truth

Model Disk Min RAM Job
qwen2.5-coder:1.5b ~1 GB 4 GB autocomplete default
qwen2.5-coder:7b ~4.7 GB 8 GB chat and edits
qwen2.5-coder:14b+ ~8 GB+ 16 GB+ not for your base Air, stop pretending

The 1.5B answers in under 400 milliseconds and stays resident next to the 7B. Start there. It covers most of a workday and all of your dignity.

❯ Choose your excuse

But the cloud model is smarter.

For the hard fifth of the job, sure. The rest is completions, renames, and boilerplate your local 7B eats cold. Stop renting a supercomputer to write getters.

But setup looks hard.

Four commands and a JSON file. You once spent a weekend theming your prompt. Fifteen minutes, once, beats an invoice forever.

But my company already pays for seats.

Then your company pays to courier its crown jewels to a vendor. Bring this page to the security review. Watch the room go quiet, then watch legal hug you.

But local is too slow.

Autocomplete lands before your finger leaves Tab. Chat takes seconds, like the cloud, minus the queue, the outage page, and the rate limit that strikes mid-demo every time, as if scheduled.

But I need the best model for everything.

No. You need the best model for the hard parts and a fast one for the rest. One-model dogma is exactly how the bill got this big.

Stop Paying Subscriptions for Shit You Can Run on Your Own Machine.

You are One Data Breach Away from Getting Fired. Local Models Don't Leak. Install the agent, pull the weights, kill the subscription.

Get OpenCode Ollama