Published: 2026-04-17
Summary

Qwen 3.6 + OpenClaw: Full Agentic Coding Locally for Free

Chapters / key moments (click to jump — plays here on the page)

Fahd Mirza connects OpenClaw to a locally-served Qwen 3.6-35B mixture-of-experts model (vLLM backend, NVIDIA H100, zero API cost) and gives it a single prompt: build a full industrial dashboard for a cement plant. The result — a working React + Vite + TypeScript app with autonomous bug detection, browser control, and Git management — demonstrates what the agentic loop looks like when there's no cloud metering constraining how far the agent runs.

Source video

"Qwen3.6-35B-A3B + OpenClaw - Agentic Coding Locally for Free" by Fahd Mirza — Watch on YouTube →

Security note

As Fahd demonstrates in the video, OpenClaw has full control of the system: it can clear browser caches, restart applications, and execute arbitrary shell commands. Run it only on machines and with accounts you are comfortable giving an agent unrestricted access to. See OpenClaw Security for setup recommendations.

Key Takeaways

OpenClaw supports custom model backends via the vLLM provider option at setup. With a locally-served model there is no API key, no per-token cost, and no rate limits — the agent runs as long as the GPU has memory and compute. This dramatically changes what autonomous agentic sessions are economically viable.
The setup process: install OpenClaw → select Quick Start → choose vLLM provider → enter local server address → specify model name → skip channels and non-essential skills → done. Fahd completes this in under 5 minutes on a fresh Ubuntu system.
The task prompt was designed to test the full agentic loop, not just code generation: "create directories, run shell commands, execute npm install, start a live server, manage Git." The question was whether Qwen 3.6, through OpenClaw's harness, could operate autonomously on a real Linux system and deliver a running application from a single instruction.
OpenClaw built the dashboard, hit a CSS rendering error, and without human instruction: detected the failure, cleared the browser cache, restarted the browser, fixed the stylesheet, and reloaded the page. This is the autonomous debugging loop — the agent treats failed renders as feedback and iterates until it passes visual inspection.
Final output: a dark-theme industrial dashboard with live sensor data simulation, alert management, production statistics, and threshold controls — plus a generated README and test files — all from one prompt with two manual follow-up corrections for layout.
Hardware requirement: NVIDIA H100 with 80GB VRAM (rented via Mass Compute in the video). Smaller models require less VRAM, but the quality of autonomous reasoning drops significantly below certain thresholds. 16GB VRAM is a reasonable floor for useful agentic work.

The Local Model Workflow

Fahd's setup: the Qwen 3.6-35B-A3B model is served via vLLM on the same Ubuntu machine where OpenClaw is installed. OpenClaw treats the local vLLM endpoint exactly as it would treat the Anthropic API — the agent loop, tool calls, file management, and terminal execution are all identical. Only the model and cost profile change.

The implication: anything you build or learn with Claude-backed OpenClaw transfers directly to a local model setup. The agent harness is the constant; the model is a swappable component. If API costs are a bottleneck on your agentic workload, this is the architecture to evaluate.

Related on OpenClawDatabase

NemoClaw Hub — NVIDIA's agent platform with local model capabilities
NemoClaw Local GPU Guide — running NemoClaw agents on local GPU
OpenClaw Security — configuring safe system access for your agent
OpenClaw Cost Reduction via NVIDIA NIM — cloud-based local GPU offloading as an alternative

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Serve the model locally first
The setup runs a local model served on an Ubuntu machine with a single <strong>H100, 80 GB VRAM</strong>, before OpenClaw is installed at all. The harness is configured against that endpoint afterwards.
Point OpenClaw at the local endpoint during setup
Through the install prompts: accept quick-start defaults, then rather than a hosted provider select the <strong>local vLLM option</strong>, give the address the server is running on, <strong>supply no API key</strong> (there is none), and name the model being served. Channels, web search and skill dependencies can all be skipped at this stage.
Verify before you trust it
<code>openclaw models list</code> confirms the model is configured and reachable. Worth doing because, as the reviewer notes, <strong>the commands change often</strong> — a week-old manual may already be wrong.
Test the full agentic loop, not just code generation
<strong>This is the reusable part of the video.</strong> The test prompt deliberately requires real tool use: create directories, run shell commands, execute a package install, start a live server and manage git. The point being made is that <em>most models can write React code</em>; the question is whether the model can operate autonomously on a real Linux system and deliver a running application from one instruction.
Iterate on failures the same way you would with a person
The first build rendered nothing. Telling the harness to fix the error was enough — it cleared the browser cache, restarted, diagnosed and fixed. A second round fixed broken stylesheets. Each cycle it closed, reloaded and tested the browser itself.

Gotchas

<strong>The harness has full control of your system</strong> — the reviewer flags this explicitly when it closes and restarts his browser unprompted. Running a local model does not reduce the harness's reach.
The GPU is rented, with a discount code in the description — a commercial relationship, disclosed.
The gateway service installed during setup is the bridge between OpenClaw core and your channels; skipping channel configuration does not skip that.
Command names move fast enough that the reviewer checks whether a week-old command still works. Expect to.

← Back to News digest · See also: OpenClaw guide