an open-source alternative to the dots, bots and muses of the world, run an office of claude code/codex like agents on your laptop, sandboxes or anywhere, uses your existing subscriptions
HTML agent-harnessagent-orchestrationagentsai-agentsautonomous-agentsclaude-codecodexcoding-agentsdesktop-appdeveloper-toolselectrongemini-cliharnessharness-engineeringlocal-firstmemorymulti-agentopencodeorchestrationtypescript
<!doctype html>
<html lang="en">
<head>
<!-- Google tag (gtag.js), Google Ads AW-18388216863. Loaded only when Do Not Track is
off, so it follows the same rule as the PostHog block on this site. -->
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){ dataLayer.push(arguments); }
if (navigator.doNotTrack !== '1') {
var gtagScript = document.createElement('script');
gtagScript.async = true;
gtagScript.src = 'https://www.googletagmanager.com/gtag/js?id=AW-18388216863';
document.head.appendChild(gtagScript);
gtag('js', new Date());
gtag('config', 'AW-18388216863');
}
</script>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Run Munder Difflin on Open Source Models: Fully Local or Through a Provider — Munder Difflin Blog</title>
<meta name="description" content="Munder Difflin can run your whole agent floor on open weight models like gpt-oss, Qwen3, DeepSeek, Llama, GLM and Kimi: fully local with Ollama, LM Studio or vLLM, or through a provider with your own key. The wiring for each engine, current as of 0.5.2." />
<link rel="canonical" href="https://munderdiffl.in/blog/run-munder-difflin-on-open-models/" />
<link rel="icon" type="image/png" href="https://munderdiffl.in/logo.png" />
<link rel="apple-touch-icon" href="https://munderdiffl.in/logo.png" />
<meta name="theme-color" content="#F5F2E8" />
<!-- Open Graph -->
<meta property="og:site_name" content="Munder Difflin Blog" />
<meta property="og:type" content="article" />
<meta property="og:title" content="Run Munder Difflin on Open Source Models: Fully Local or Through a Provider — Munder Difflin Blog" />
<meta property="og:description" content="Munder Difflin can run your whole agent floor on open weight models like gpt-oss, Qwen3, DeepSeek, Llama, GLM and Kimi: fully local with Ollama, LM Studio or vLLM, or through a provider with your own key. The wiring for each engine, current as of 0.5.2." />
<meta property="og:url" content="https://munderdiffl.in/blog/run-munder-difflin-on-open-models/" />
<meta property="og:image" content="https://munderdiffl.in/blog/assets/media/run-munder-difflin-on-open-models/hero.png" />
<meta property="og:image:width" content="1600" />
<meta property="og:image:height" content="904" />
<meta property="og:image:alt" content="Illustration: Run Munder Difflin on Open-Source Models — Fully Local or via a Third-Party Provider" />
<meta property="og:locale" content="en_US" />
<meta property="article:published_time" content="2026-06-22T00:00:00.000Z" />
<meta property="article:author" content="Chaitanya Giri" />
<meta property="article:tag" content="Guides" />
<meta property="article:tag" content="Local-First" />
<meta property="article:tag" content="Open Source" />
<meta property="article:tag" content="CLI Agents" />
<meta property="article:tag" content="Tutorial" />
<!-- Twitter -->
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="Run Munder Difflin on Open Source Models: Fully Local or Through a Provider — Munder Difflin Blog" />
<meta name="twitter:description" content="Munder Difflin can run your whole agent floor on open weight models like gpt-oss, Qwen3, DeepSeek, Llama, GLM and Kimi: fully local with Ollama, LM Studio or vLLM, or through a provider with your own key. The wiring for each engine, current as of 0.5.2." />
<meta name="twitter:image" content="https://munderdiffl.in/blog/assets/media/run-munder-difflin-on-open-models/hero.png" />
<meta name="twitter:image:alt" content="Illustration: Run Munder Difflin on Open-Source Models — Fully Local or via a Third-Party Provider" />
<link rel="alternate" type="application/atom+xml" title="Munder Difflin Blog" href="https://munderdiffl.in/blog/feed.xml" />
<link rel="preconnect" href="https://fonts.googleapis.com" />
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
<link href="https://fonts.googleapis.com/css2?family=Fraunces:opsz,[email protected],500;9..144,600;9..144,700;9..144,900&family=Source+Serif+4:opsz,[email protected],400;8..60,600;8..60,700&family=Space+Grotesk:wght@500;600;700&family=JetBrains+Mono:wght@400;600&display=swap" rel="stylesheet" />
<link rel="stylesheet" href="/blog/assets/blog-press.css" />
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "BlogPosting",
"headline": "Run Munder Difflin on Open Source Models: Fully Local or Through a Provider",
"description": "Munder Difflin can run your whole agent floor on open weight models like gpt-oss, Qwen3, DeepSeek, Llama, GLM and Kimi: fully local with Ollama, LM Studio or vLLM, or through a provider with your own key. The wiring for each engine, current as of 0.5.2.",
"image": ["https://munderdiffl.in/blog/assets/media/run-munder-difflin-on-open-models/hero.png"],
"datePublished": "2026-06-22T00:00:00.000Z",
"dateModified": "2026-09-10T00:00:00.000Z",
"author": { "@type": "Person", "name": "Chaitanya Giri" },
"publisher": {
"@type": "Organization",
"name": "Munder Difflin",
"logo": { "@type": "ImageObject", "url": "https://munderdiffl.in/logo.png" }
},
"mainEntityOfPage": { "@type": "WebPage", "@id": "https://munderdiffl.in/blog/run-munder-difflin-on-open-models/" },
"keywords": "run ai agents on open source models, local llm coding agent, ollama coding agent, openrouter coding agent, gpt-oss, byok open models, opencode crush pi, pi models.json ollama",
"articleSection": "Guides",
"inLanguage": "en-US",
"url": "https://munderdiffl.in/blog/run-munder-difflin-on-open-models/"
}
</script>
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "BreadcrumbList",
"itemListElement": [
{ "@type": "ListItem", "position": 1, "name": "Home", "item": "https://munderdiffl.in/" },
{ "@type": "ListItem", "position": 2, "name": "Blog", "item": "https://munderdiffl.in/blog/" },
{ "@type": "ListItem", "position": 3, "name": "Run Munder Difflin on Open Source Models: Fully Local or Through a Provider", "item": "https://munderdiffl.in/blog/run-munder-difflin-on-open-models/" }
]
}
</script>
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{ "@type": "Question", "name": "Can Munder Difflin run entirely on open source models?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. OpenCode, Crush, Qwen and Pi can all run open weight models, either on your own machine or through a provider with your own API key. You can put open models in every seat on the floor, including Michael's." } },
{ "@type": "Question", "name": "What is the difference between running local and using a provider?", "acceptedAnswer": { "@type": "Answer", "text": "Local means Ollama, LM Studio or vLLM runs the weights on your machine: private, no per token bill, and limited by your memory. A provider such as OpenRouter or Groq hosts the same open weights on its hardware and bills your own key per token, so you can reach models far too big for a laptop." } },
{ "@type": "Question", "name": "Which open model should I give the orchestrator?", "acceptedAnswer": { "@type": "Answer", "text": "A strong one, because Michael does the reasoning and the long context coordination. Locally that means gpt-oss 120B or Llama 3.3 70B on a machine with 64 to 96 GB of memory. Through a provider, DeepSeek V4 Flash or Kimi K2.6 on OpenRouter. Models of 8B and under make fine workers and thin orchestrators." } },
{ "@type": "Question", "name": "Does each engine need a different model name?", "acceptedAnswer": { "@type": "Answer", "text": "The model id stays the same and only the prefix changes. A local model is local/<tag> on OpenCode and ollama/<tag> on Crush and Pi, and a provider model carries the provider first, like openrouter/openai/gpt-oss-120b." } },
{ "@type": "Question", "name": "How do I run Pi on a local model?", "acceptedAnswer": { "@type": "Answer", "text": "Add your local server to Pi's own config file, ~/.pi/agent/models.json. Munder Difflin copies that file into every Pi agent it starts, so the models you define there are available to your Pi agents. The Pi base URL field in the app stays reserved." } }
]
}
</script>
<script>
// Anonymous blog analytics (PostHog) — same posture as the main site and the
// app's TELEMETRY.md: public write-only token ('' = inert), anonymous-only
// (person_profiles 'never'), no autocapture/recording, DNT respected.
(function () {
var PH_KEY = 'phc_qJzsokwcbtTGHemnvywR3nSqivNNgMjHeNVEWkZNydFB';
if (!PH_KEY || navigator.doNotTrack === '1') return;
var s = document.createElement('script');
s.async = true;
s.src = 'https://us-assets.i.posthog.com/static/array.js';
s.onload = function () {
if (!window.posthog) return;
window.posthog.init(PH_KEY, {
api_host: 'https://us.i.posthog.com',
person_profiles: 'never',
autocapture: false,
before_send: function (event) {
// PostHog fills $ip from the connection unless the event carries one.
// Null it here and disable the GeoIP lookup so no IP, city, postal
// code or lat/long is ever derived or stored for this page view.
if (event && event.properties) {
event.properties.$ip = null;
event.properties.$geoip_disable = true;
}
return event;
},
disable_session_recording: true,
respect_dnt: true
});
};
document.head.appendChild(s);
})();
</script>
</head>
<body class="theme-press">
<a class="skip-link" href="#main">Skip to content</a>
<nav id="nav">
<a class="brand" href="https://munderdiffl.in" aria-label="Munder Difflin home">
<img class="brand-logo" src="https://munderdiffl.in/logo.png" width="1318" height="840" alt="Munder Difflin, Inc. — Multi-Agent Harness" />
</a>
<span class="spacer"></span>
<div class="links">
<a href="/blog/">Latest</a>
<a href="/blog/topics/">Topics</a>
<a href="/blog/about/">About</a>
<a href="https://munderdiffl.in">↗ Product</a>
</div>
<div class="nav-actions">
<a class="btn sm" href="https://munderdiffl.in/blog/feed.xml" aria-label="RSS feed">⤵ RSS</a>
<a class="btn sm primary" href="https://munderdiffl.in/#install">⤓ Download</a>
</div>
</nav>
<main id="main">
<article class="t-guides">
<header class="post-head wrap">
<nav class="crumbs" aria-label="Breadcrumb">
<a href="/blog/">Blog</a>
<span>/</span>
<a href="/blog/topics/guides/">Guides</a>
</nav>
<h1>Run Munder Difflin on Open Source Models: Fully Local or Through a Provider</h1>
<p class="dek">Munder Difflin can run your whole agent floor on open weight models like gpt-oss, Qwen3, DeepSeek, Llama, GLM and Kimi: fully local with Ollama, LM Studio or vLLM, or through a provider with your own key. The wiring for each engine, current as of 0.5.2.</p>
<div class="byline">
<span class="av" aria-hidden="true">CG</span>
<strong>Chaitanya Giri</strong>
<span class="dot" aria-hidden="true">·</span>
<time datetime="2026-06-22">Jun 22, 2026</time>
<span class="dot" aria-hidden="true">·</span>
<span>6 min read</span>
<span class="tag kind">Guides</span>
</div>
</header>
<div class="hero-wrap wrap"><figure class="fig fig-hero"><img src="/blog/assets/media/run-munder-difflin-on-open-models/hero.png" alt="Illustration: Run Munder Difflin on Open-Source Models — Fully Local or via a Third-Party Provider" loading="lazy" decoding="async" /></figure></div>
<div class="post-wrap wrap">
<div class="prose">
<div class="callout tldr"><span class="ic">TL;DR</span><p><strong>Munder Difflin can run on open models
end to end.</strong> Two routes: <strong>fully local</strong> with Ollama, LM Studio or vLLM (private, no per
token bill, limited by your memory), or a <strong>provider</strong> such as OpenRouter or Groq (their hardware,
your key, much bigger models). Four engines do the wiring: <strong>OpenCode</strong>, <strong>Crush</strong> and
<strong>Qwen</strong> take a local base URL in <strong>Settings → AI Engines</strong>, and <strong>Pi</strong> reads
your own <code>~/.pi/agent/models.json</code>. Keys go in the same panel and are stored write only.</p></div>
<p>Munder Difflin started out wrapping the closed frontier CLIs. Today it supports twelve, and several of them will
point at any model you like. That means a whole office of agents can run on models whose weights anyone can download.</p>
<p>There are two honest ways to do it, and they trade off differently. This guide walks through both, then gives the exact
wiring for each engine, checked against Munder Difflin 0.5.2 on 10 September 2026. (For the why, see
<a href="/blog/why-local-first-matters-for-ai-agents/">why local first matters for AI agents</a>.)</p>
<h2 id="should-you-run-open-models-locally-or-through-a-provider" tabindex="-1">Should you run open models locally or through a provider? <a class="anchor" href="#should-you-run-open-models-locally-or-through-a-provider" aria-hidden="true">#</a></h2>
<p>Local if privacy and a fixed cost matter and your machine can hold the model. A provider if you want the biggest models
or have no spare hardware.</p>
<div class="table-scroll">
<table>
<thead>
<tr>
<th></th>
<th>Fully local</th>
<th>Through a provider</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Runs on</strong></td>
<td>Your machine (Ollama, LM Studio, vLLM)</td>
<td>Their GPUs (OpenRouter, Groq and others)</td>
</tr>
<tr>
<td><strong>Cost</strong></td>
<td>Electricity. No per token bill.</td>
<td>Per token, billed to your own key.</td>
</tr>
<tr>
<td><strong>Privacy</strong></td>
<td>Prompts never leave the machine.</td>
<td>Prompts go to the provider.</td>
</tr>
<tr>
<td><strong>Ceiling</strong></td>
<td>Your memory: roughly 8B to 70B on a Mac, 120B on a very big one.</td>
<td>Open models with hundreds of billions of parameters.</td>
</tr>
<tr>
<td><strong>Setup</strong></td>
<td>Pull a model and point the engine at localhost.</td>
<td>Paste one key.</td>
</tr>
<tr>
<td><strong>Best for</strong></td>
<td>Private work, always on floors, fixed cost.</td>
<td>Top quality, bursty use, no local hardware.</td>
</tr>
</tbody>
</table>
</div>
<p>You do not have to choose once for the whole floor. Engine and model are set per agent, so a strong provider model can sit
in Michael’s seat while cheap local workers handle the routine majority. That is
<a href="/blog/do-more-with-less-model-routing/">capability routing</a> with open weights at both ends.</p>
<h2 id="which-engines-can-run-open-models" tabindex="-1">Which engines can run open models? <a class="anchor" href="#which-engines-can-run-open-models" aria-hidden="true">#</a></h2>
<p>Four, and each one wires it a little differently.</p>
<div class="table-scroll">
<table>
<thead>
<tr>
<th>Engine</th>
<th>Where the local server goes</th>
<th>What the app does with it</th>
<th>Local model slug</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>OpenCode</strong></td>
<td>Base URL in Settings → AI Engines</td>
<td>Injects it as a local OpenAI compatible provider named <code>local</code></td>
<td><code>local/<tag></code></td>
</tr>
<tr>
<td><strong>Crush</strong></td>
<td>Base URL in Settings → AI Engines</td>
<td>Uses it as the upstream of the local proxy Crush runs through</td>
<td><code>ollama/<tag></code></td>
</tr>
<tr>
<td><strong>Qwen</strong></td>
<td>Base URL and default model in Settings → AI Engines</td>
<td>The same proxy approach as Crush</td>
<td>the default model you set</td>
</tr>
<tr>
<td><strong>Pi</strong></td>
<td>Your own <code>~/.pi/agent/models.json</code></td>
<td>Copies that file into every Pi agent it starts</td>
<td><code>ollama/<tag></code></td>
</tr>
</tbody>
</table>
</div>
<p>The usual endpoints: Ollama at <code>http://localhost:11434/v1</code>, LM Studio at <code>http://localhost:1234/v1</code>, and vLLM wherever you
expose it, often <code>:8000/v1</code>.</p>
<figure class="fig fig-inline"><img src="/blog/assets/media/run-munder-difflin-on-open-models/note-1.png" alt="Hand-drawn sketch from “Run Munder Difflin on Open-Source Models — Fully Local or via a Third-Party Provider”" loading="lazy" decoding="async" /><figcaption>Same model id everywhere. Only the prefix changes: local/ on OpenCode, ollama/ on Crush and Pi, and the provider name when you use a key.</figcaption></figure>
<h2 id="how-do-you-run-a-model-fully-locally" tabindex="-1">How do you run a model fully locally? <a class="anchor" href="#how-do-you-run-a-model-fully-locally" aria-hidden="true">#</a></h2>
<p>Pull a model, tell the engine where it lives, and pick it for an agent.</p>
<p>Want a local model behind Claude Code itself? That route is covered in <a href="/blog/how-to-connect-ollama-to-claude-code/">how to connect Ollama to Claude Code</a>. Hosted models instead? See <a href="/blog/claude-code-router/">Claude Code Router</a>. A newer open weight model to watch is <a href="/blog/reflection-beam/">Reflection AI’s Beam</a>; its weights were not out when we checked on 6 Oct 2026.</p>
<p><strong>1. Pull a model.</strong> With <a href="https://ollama.com">Ollama</a> installed, grab one sized to your memory:</p>
<pre class="language-bash"><code class="language-bash">ollama pull gpt-oss:20b <span class="token comment"># about 14 GB, fits a 16 GB Mac</span>
ollama pull qwen3:30b-a3b <span class="token comment"># about 19 GB, a fast generalist for 32 GB</span>
ollama pull deepseek-r1:32b <span class="token comment"># about 20 GB, strong reasoning for 32 GB</span>
ollama serve <span class="token comment"># serves the OpenAI compatible API on port 11434</span></code></pre>
<p>LM Studio works the same way: load a model in the app and start its local server on port 1234.</p>
<p><strong>2. Point the engine at it.</strong> For OpenCode, Crush or Qwen, open <strong>Settings → AI Engines</strong> and set that engine’s local base
URL, for example <code>http://localhost:11434/v1</code>. Local servers need no key. For Pi, add the server to <code>~/.pi/agent/models.json</code>
instead:</p>
<pre class="language-json"><code class="language-json"><span class="token punctuation">{</span>
<span class="token property">"providers"</span><span class="token operator">:</span> <span class="token punctuation">{</span>
<span class="token property">"ollama"</span><span class="token operator">:</span> <span class="token punctuation">{</span>
<span class="token property">"baseUrl"</span><span class="token operator">:</span> <span class="token string">"http://localhost:11434/v1"</span><span class="token punctuation">,</span>
<span class="token property">"api"</span><span class="token operator">:</span> <span class="token string">"openai-completions"</span><span class="token punctuation">,</span>
<span class="token property">"apiKey"</span><span class="token operator">:</span> <span class="token string">"ollama"</span><span class="token punctuation">,</span>
<span class="token property">"models"</span><span class="token operator">:</span> <span class="token punctuation">[</span><span class="token punctuation">{</span> <span class="token property">"id"</span><span class="token operator">:</span> <span class="token string">"gpt-oss:20b"</span> <span class="token punctuation">}</span><span class="token punctuation">]</span>
<span class="token punctuation">}</span>
<span class="token punctuation">}</span>
<span class="token punctuation">}</span></code></pre>
<p>Munder Difflin copies that file into each Pi agent when it starts, so start the Pi agent again after you edit it.</p>
<p><strong>3. Hire an agent on that model.</strong> In <strong>Add agent</strong>, choose the engine and pick one of the open model quick picks, or type
the slug yourself: <code>local/gpt-oss:20b</code> on OpenCode, <code>ollama/gpt-oss:20b</code> on Crush and Pi. Keep the colon in the tag. That
agent now runs entirely on your hardware.</p>
<p>Which local model? These are the app’s local quick picks, by memory:</p>
<div class="table-scroll">
<table>
<thead>
<tr>
<th>Model</th>
<th>Ollama tag</th>
<th>Memory</th>
<th>Good for</th>
</tr>
</thead>
<tbody>
<tr>
<td>gpt-oss 20B</td>
<td><code>gpt-oss:20b</code></td>
<td>16 GB</td>
<td>The smallest capable default</td>
</tr>
<tr>
<td>Mistral Small 24B</td>
<td><code>mistral-small:24b</code></td>
<td>16 to 32 GB</td>
<td>A light generalist</td>
</tr>
<tr>
<td>Qwen3 30B A3B</td>
<td><code>qwen3:30b-a3b</code></td>
<td>32 GB</td>
<td>A fast generalist</td>
</tr>
<tr>
<td>Qwen3 Coder 30B</td>
<td><code>qwen3-coder:30b</code></td>
<td>32 GB</td>
<td>Coding</td>
</tr>
<tr>
<td>DeepSeek R1 32B</td>
<td><code>deepseek-r1:32b</code></td>
<td>32 GB</td>
<td>Reasoning</td>
</tr>
<tr>
<td>GLM 4.7 Flash</td>
<td><code>glm-4.7-flash</code></td>
<td>32 GB</td>
<td>The GLM that fits on a Mac</td>
</tr>
<tr>
<td>Llama 3.3 70B</td>
<td><code>llama3.3:70b</code></td>
<td>64 GB</td>
<td>A bigger generalist</td>
</tr>
<tr>
<td>gpt-oss 120B</td>
<td><code>gpt-oss:120b</code></td>
<td>96 GB</td>
<td>The top local pick</td>
</tr>
</tbody>
</table>
</div>
<p>The headline open flagships, like DeepSeek V4, Kimi K2.6 and Qwen3 235B, are server class. No consumer Mac holds them, so
use a provider for those. Sizing a model to your memory is the whole subject of the
<a href="/blog/run-munder-difflin-on-a-mac-mini/">Mac mini guide</a>.</p>
<h2 id="how-do-you-run-open-models-through-a-provider" tabindex="-1">How do you run open models through a provider? <a class="anchor" href="#how-do-you-run-open-models-through-a-provider" aria-hidden="true">#</a></h2>
<p>Same open weights, someone else’s GPUs, your own key.</p>
<p><strong>1. Get a key.</strong> <a href="https://openrouter.ai">OpenRouter</a> is the easiest start, with one key for a wide catalog.
<a href="https://groq.com">Groq</a> is quick for the models it carries.</p>
<p><strong>2. Paste it into Settings → AI Engines.</strong> There are key fields for Anthropic, OpenAI, Google Gemini, OpenRouter and Groq.
Keys are write only: the app stores a key but never shows it back, and it reaches an agent as the right environment variable
only when that agent starts.</p>
<p><strong>3. Pick a model.</strong> In Add agent, choose OpenCode, Crush or Pi and a provider hosted quick pick:</p>
<div class="table-scroll">
<table>
<thead>
<tr>
<th>Model</th>
<th>Route</th>
<th>Slug</th>
<th>Key</th>
</tr>
</thead>
<tbody>
<tr>
<td>gpt-oss 120B</td>
<td>Groq</td>
<td><code>groq/openai/gpt-oss-120b</code></td>
<td><code>GROQ_API_KEY</code></td>
</tr>
<tr>
<td>Llama 3.3 70B</td>
<td>Groq</td>
<td><code>groq/llama-3.3-70b-versatile</code></td>
<td><code>GROQ_API_KEY</code></td>
</tr>
<tr>
<td>DeepSeek V4 Flash</td>
<td>OpenRouter</td>
<td><code>openrouter/deepseek/deepseek-v4-flash</code></td>
<td><code>OPENROUTER_API_KEY</code></td>
</tr>
<tr>
<td>GLM 4.6</td>
<td>OpenRouter</td>
<td><code>openrouter/z-ai/glm-4.6</code></td>
<td><code>OPENROUTER_API_KEY</code></td>
</tr>
<tr>
<td>Kimi K2.6</td>
<td>OpenRouter</td>
<td><code>openrouter/moonshotai/kimi-k2.6</code></td>
<td><code>OPENROUTER_API_KEY</code></td>
</tr>
<tr>
<td>Qwen3 Coder 480B</td>
<td>OpenRouter</td>
<td><code>openrouter/qwen/qwen3-coder</code></td>
<td><code>OPENROUTER_API_KEY</code></td>
</tr>
<tr>
<td>Qwen3 235B</td>
<td>OpenRouter</td>
<td><code>openrouter/qwen/qwen3-235b-a22b-2507</code></td>
<td><code>OPENROUTER_API_KEY</code></td>
</tr>
<tr>
<td>gpt-oss 120B</td>
<td>OpenRouter</td>
<td><code>openrouter/openai/gpt-oss-120b</code></td>
<td><code>OPENROUTER_API_KEY</code></td>
</tr>
</tbody>
</table>
</div>
<figure class="fig fig-inline"><img src="/blog/assets/media/run-munder-difflin-on-open-models/note-2.png" alt="Hand-drawn sketch from “Run Munder Difflin on Open-Source Models — Fully Local or via a Third-Party Provider”" loading="lazy" decoding="async" /><figcaption>One key, the whole catalog: paste it once, and it reaches an agent only when that agent starts.</figcaption></figure>
<h2 id="what-should-go-in-michaels-seat" tabindex="-1">What should go in Michael’s seat? <a class="anchor" href="#what-should-go-in-michaels-seat" aria-hidden="true">#</a></h2>
<p>A strong model. Michael does the reasoning, holds the long context and decides who does what. Locally, that is <code>gpt-oss:120b</code>
or <code>llama3.3:70b</code> on 64 to 96 GB of memory. Through a provider, DeepSeek V4 Flash or Kimi K2.6. Models under 8B make good
workers and thin orchestrators.</p>
<h2 id="what-if-it-does-not-work" tabindex="-1">What if it does not work? <a class="anchor" href="#what-if-it-does-not-work" aria-hidden="true">#</a></h2>
<ul>
<li><strong>The model name is rejected.</strong> Check the prefix: <code>local/</code> on OpenCode, <code>ollama/</code> on Crush and Pi. Keep the colon in the
Ollama tag.</li>
<li><strong>The agent cannot reach the server.</strong> Make sure <code>ollama serve</code> or LM Studio’s server is running, and that the base URL
ends in <code>/v1</code>.</li>
<li><strong>A Pi agent cannot see your model.</strong> Check that <code>~/.pi/agent/models.json</code> is valid JSON, then start the Pi agent again so
the app copies the new file.</li>
<li><strong>An engine shows as missing.</strong> Settings → Prerequisites lists which engine commands the app can actually find.</li>
</ul>
<h2 id="the-bottom-line" tabindex="-1">The bottom line <a class="anchor" href="#the-bottom-line" aria-hidden="true">#</a></h2>
<p>Open weights turn Munder Difflin from a harness for a few vendors’ CLIs into a harness for the whole open ecosystem. Go local
when privacy and a fixed cost matter, use a provider when you want the biggest models, and mix both across your floor, agent
by agent.</p>
<p><a href="https://munderdiffl.in/">Download Munder Difflin</a>, free and open source, and point your favourite open model at it. On a Mac
and want the sizing walkthrough? Read the <a href="/blog/run-munder-difflin-on-a-mac-mini/">Mac mini guide</a>.</p>
<section class="faq" aria-label="Frequently asked questions">
<h2 id="faq">FAQ</h2>
<div class="qa">
<p class="q">Can Munder Difflin run entirely on open source models?</p>
<p class="a">Yes. OpenCode, Crush, Qwen and Pi can all run open weight models, either on your own machine or through a provider with your own API key. You can put open models in every seat on the floor, including Michael's.</p>
</div>
<div class="qa">
<p class="q">What is the difference between running local and using a provider?</p>
<p class="a">Local means Ollama, LM Studio or vLLM runs the weights on your machine: private, no per token bill, and limited by your memory. A provider such as OpenRouter or Groq hosts the same open weights on its hardware and bills your own key per token, so you can reach models far too big for a laptop.</p>
</div>
<div class="qa">
<p class="q">Which open model should I give the orchestrator?</p>
<p class="a">A strong one, because Michael does the reasoning and the long context coordination. Locally that means gpt-oss 120B or Llama 3.3 70B on a machine with 64 to 96 GB of memory. Through a provider, DeepSeek V4 Flash or Kimi K2.6 on OpenRouter. Models of 8B and under make fine workers and thin orchestrators.</p>
</div>
<div class="qa">
<p class="q">Does each engine need a different model name?</p>
<p class="a">The model id stays the same and only the prefix changes. A local model is local/<tag> on OpenCode and ollama/<tag> on Crush and Pi, and a provider model carries the provider first, like openrouter/openai/gpt-oss-120b.</p>
</div>
<div class="qa">
<p class="q">How do I run Pi on a local model?</p>
<p class="a">Add your local server to Pi's own config file, ~/.pi/agent/models.json. Munder Difflin copies that file into every Pi agent it starts, so the models you define there are available to your Pi agents. The Pi base URL field in the app stays reserved.</p>
</div>
</section>
<hr />
<div class="share">
<span class="lbl">Tags</span>
<a class="tag" href="/blog/tags/guides/">Guides</a><a class="tag" href="/blog/tags/local-first/">Local-First</a><a class="tag" href="/blog/tags/open-source/">Open Source</a><a class="tag" href="/blog/tags/cli-agents/">CLI Agents</a><a class="tag" href="/blog/tags/tutorial/">Tutorial</a>
</div>
</div>
<aside class="toc-rail" aria-label="Table of contents">
<div class="tt">On this page</div>
<ol>
<li class="lvl-2"><a href="#should-you-run-open-models-locally-or-through-a-provider">Should you run open models locally or through a provider?</a></li><li class="lvl-2"><a href="#which-engines-can-run-open-models">Which engines can run open models?</a></li><li class="lvl-2"><a href="#how-do-you-run-a-model-fully-locally">How do you run a model fully locally?</a></li><li class="lvl-2"><a href="#how-do-you-run-open-models-through-a-provider">How do you run open models through a provider?</a></li><li class="lvl-2"><a href="#what-should-go-in-michaels-seat">What should go in Michael’s seat?</a></li><li class="lvl-2"><a href="#what-if-it-does-not-work">What if it does not work?</a></li><li class="lvl-2"><a href="#the-bottom-line">The bottom line</a></li>
</ol>
</aside>
</div>
<footer class="post-foot wrap"><div class="prevnext">
<a class="pn prev" href="/blog/launching-munder-difflin-v0-3-0/"><span class="k">← Older</span><span class="t">Launching Munder Difflin v0.3.0: Selectable Engines, Integrations & Slack-Spawned Workers</span></a>
<a class="pn next" href="/blog/run-munder-difflin-on-a-mac-mini/"><span class="k">Newer →</span><span class="t">Run Munder Difflin Locally on a Mac Mini</span></a>
</div>
<section class="related">
<h2>Related in Guides</h2>
<div class="post-grid">
<article class="card t-guides">
<a class="thumb t-guides" href="/blog/claude-md-best-practices/" aria-hidden="true" tabindex="-1"><img src="/blog/assets/media/claude-md-best-practices/hero.png" alt="" loading="lazy" decoding="async" /></a>
<div class="body">
<div class="meta">
<time datetime="2026-10-26">Oct 26, 2026</time>
<span class="dot" aria-hidden="true"></span>
<span>4 min</span>
</div>
<h3><a href="/blog/claude-md-best-practices/">CLAUDE.md Best Practices: What to Write, Where It Lives, How Long (Oct 2026)</a></h3>
<p class="dek">CLAUDE.md best practices from Anthropic's docs, checked 26 Oct 2026: what to put in it, where each file lives, the 200 line target and how AGENTS.md loads.</p>
<div class="foot">
<span class="tag kind">Guides</span>
<a class="more" href="/blog/claude-md-best-practices/" aria-label="Read: CLAUDE.md Best Practices: What to Write, Where It Lives, How Long (Oct 2026)">Read →</a>
</div>
</div>
</article>
<article class="card t-guides">
<a class="thumb t-guides" href="/blog/codex-computer-use/" aria-hidden="true" tabindex="-1"><img src="/blog/assets/media/codex-computer-use/hero.png" alt="" loading="lazy" decoding="async" /></a>
<div class="body">
<div class="meta">
<time datetime="2026-10-26">Oct 26, 2026</time>
<span class="dot" aria-hidden="true"></span>
<span>4 min</span>
</div>
<h3><a href="/blog/codex-computer-use/">Codex Computer Use: How to Turn It On, Plans and Limits (Oct 2026)</a></h3>
<p class="dek">Codex computer use lets the ChatGPT desktop app see and click apps on macOS and Windows. Setup, which plans get it, locked use and how Claude's compares.</p>
<div class="foot">
<span class="tag kind">Guides</span>
<a class="more" href="/blog/codex-computer-use/" aria-label="Read: Codex Computer Use: How to Turn It On, Plans and Limits (Oct 2026)">Read →</a>
</div>
</div>
</article>
<article class="card t-guides">
<a class="thumb t-guides" href="/blog/cursor-pricing/" aria-hidden="true" tabindex="-1"><img src="/blog/assets/media/cursor-pricing/hero.png" alt="" loading="lazy" decoding="async" /></a>
<div class="body">
<div class="meta">
<time datetime="2026-10-26">Oct 26, 2026</time>
<span class="dot" aria-hidden="true"></span>
<span>5 min</span>
</div>
<h3><a href="/blog/cursor-pricing/">Cursor Pricing (Oct 2026): Every Plan After the 12 Oct Change</a></h3>
<p class="dek">Cursor pricing as of 26 Oct 2026: Free, Plus $8, Premium $30, Super $100, Ultra $200, Teams seats, what changed on 12 Oct and what old plans keep.</p>
<div class="foot">
<span class="tag kind">Guides</span>
<a class="more" href="/blog/cursor-pricing/" aria-label="Read: Cursor Pricing (Oct 2026): Every Plan After the 12 Oct Change">Read →</a>
</div>
</div>
</article>
</div>
</section>
</footer>
</article>
</main>
<footer class="site-foot">
<div class="wrap">
<div class="foot-top">
<a class="foot-brand brand" href="https://munderdiffl.in" aria-label="Munder Difflin home">
<img class="brand-logo" src="https://munderdiffl.in/logo.png" width="1318" height="840" alt="Munder Difflin, Inc. — Multi-Agent Harness" />
</a>
<div class="foot-links">
<a href="/blog/">Latest</a>
<a href="/blog/topics/">Topics</a>
<a href="/blog/about/">About</a>
<a href="https://munderdiffl.in/blog/feed.xml">RSS</a>
<a href="https://github.com/chaitanyagiri/munder-difflin">GitHub</a>
<a href="https://munderdiffl.in">munderdiffl.in ↗</a>
</div>
</div>
<div class="foot-meta">
© 2026 Munder Difflin · Local-first · MIT-licensed ·
Built on <a href="https://munderdiffl.in">Electron, React, Pixi.js, xterm.js & node-pty</a>.
Run an office of agents while you sleep.
</div>
</div>
</footer>
<script>
// sticky-nav border on scroll
(function () {
var nav = document.getElementById('nav');
if (!nav) return;
var onScroll = function () { nav.classList.toggle('scrolled', window.scrollY > 8); };
onScroll(); window.addEventListener('scroll', onScroll, { passive: true });
})();
// YouTube click-to-load: thumbnail button → youtube-nocookie iframe
(function () {
document.addEventListener('click', function (e) {
var btn = e.target.closest('.yt-load');
if (!btn) return;
var box = btn.closest('.yt');
var id = box && box.getAttribute('data-yt');
// Only an 11 character YouTube video id may reach the iframe URL.
if (!id || !/^[A-Za-z0-9_-]{11}$/.test(id)) return;
var frame = document.createElement('iframe');
frame.src = 'https://www.youtube-nocookie.com/embed/' + encodeURIComponent(id) + '?autoplay=1';
frame.title = box.getAttribute('data-title') || 'YouTube video';
frame.allow = 'accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share';
frame.setAttribute('allowfullscreen', '');
box.replaceChildren(frame);
});
})();
</script>
</body>
</html>