Quick Overview
This technical walkthrough by Nate Herk demonstrates how to configure the Claude Code agent harness to run entirely on free models hosted on OpenRouter. The video explains the required configuration changes in VS Code and evaluates the real-world performance of an anonymous stealth reasoning model across web development and analytics tasks.
Key Points
- 1.Claude Code can be configured to use free external models hosted on OpenRouter by editing environment variables in the local configuration file.
- 2.The desktop version of Claude Code overrides custom model settings, requiring users to run Claude Code through an IDE terminal like VS Code.
- 3.Third-party models such as the stealth Ox Alpha model retain access to Claude Code tool harnesses, sub-agents, memory systems, and custom skills.
- 4.High-token generation requests using free OpenRouter models frequently encounter upstream idle timeout errors unless tasks are broken down into smaller chunks.
- 5.While free stealth models can successfully complete complex web design and analytical tasks, their processing speed is substantially slower than native Anthropic models.
Summary
The video demonstrates how to run the Claude Code developer harness with free external language models by connecting it to OpenRouter. Nate Herk begins by showing active Claude Code terminal sessions executing tasks under an anonymous free model called stealth/ox-alpha rather than native Anthropic Claude models. He explains that while the standard Claude Code desktop application overrides external routing and forces default Anthropic models, developers can bypass this restriction by running Claude Code inside a code editor like VS Code or an independent command terminal.
To establish the integration, users navigate to OpenRouter, generate an account API key, and review available free models. Within their project directory, users open the hidden settings configuration file and update the environment block. This requires setting the base URL to the OpenRouter API endpoint, supplying the OpenRouter API key in the authentication token field, and assigning all model alias fields, including Sonnet, Opus, Haiku, and sub-agent defaults, to the target model string. Herk also points out alternative options, such as GLM 5.2 Free or the dynamic Free Models Router, which automatically cycles through available free endpoints.
Herk tests the system by assigning a full front-end web development task to build a branded landing page and interactive e-commerce product catalog. The stealth model accesses the project folder, reads brand styling guidelines, and executes automated verification checks across the codebase. During large file generations, the model repeatedly hits upstream idle timeout errors caused by proxy limitations. Herk resolves this by prompting the agent to write files in smaller modular chunks and provide handoff messages, allowing the site build and automated browser tests to finish successfully over a six-hour duration.
Additional experiments demonstrate that the model successfully integrates with Claude Code memory, workspace rules, and custom automation skills. In a data analytics test, the model investigates historical scripts across neighboring directories, fetches YouTube analytics data for 469 videos, and builds a formatted multi-sheet performance report. In another test, it parses 166 community discussions to classify support topics. Herk concludes that while free third-party models work effectively for knowledge tasks, their slow execution speed makes them best suited as budget-friendly fallbacks when standard API credits run out.
Routing Claude Code Through OpenRouter
Nate Herk shows that Claude Code can be pointed away from paid Anthropic endpoints to OpenRouter, an aggregator hosting hundreds of models including free options. Because the standalone Claude Code desktop app enforces standard Anthropic model usage, the custom configuration must be run through a terminal or an IDE such as VS Code. Setup requires obtaining an OpenRouter API key and modifying the environment settings in the local Claude configuration file.
Configuring Model Settings and the Stealth Provider
Users configure the settings.local.json file by defining the base URL for OpenRouter, replacing the Anthropic authentication token with their OpenRouter API key, and updating model target fields to specific model identifiers like stealth/ox-alpha or z-ai/glm-5.2:free. The video notes that models labeled under the stealth provider operate under anonymous terms of service, which users should consider when handling sensitive data.
Testing Coding Capabilities on a Web Build
To evaluate coding ability, Herk prompts the stealth model to generate a multi-page product landing page and functional e-commerce shopping cart using local brand assets. The model successfully navigates local directories, creates web files, and executes browser-based verification routines. However, heavy file writing triggers upstream idle timeout errors, requiring prompts to break down outputs into smaller chunks and extending total build time to six hours.
Evaluating Data Analysis and Knowledge Work
The model is tested on knowledge-work workflows, including generating a multi-tab Google Sheet report analyzing YouTube channel metrics and categorizing community support posts. The model successfully accesses workspace skills, scans historical project files, and extracts relevant analytical trends. While the final analytical outputs match the quality of standard models, execution times remain considerably longer.
The Bottom Line
The video establishes that developers can successfully power the Claude Code harness using free models via OpenRouter configuration files. It proves that tools, sub-agents, and workspace skills remain functional across external models without losing context awareness. However, severe processing speed limitations and proxy timeout errors leave free models practical primarily for low-urgency tasks or credit exhausted fallback scenarios.
FAQ
What is the stealth Ox Alpha model and how can it be used in Claude Code?
Stealth Ox Alpha is a free reasoning model hosted on OpenRouter that can be connected to Claude Code by replacing the default Anthropic API endpoints and keys in the local settings file.
Why does the desktop version of Claude Code prevent users from switching to third-party models?
The desktop application hardcodes configuration overrides that force sessions to use native Anthropic Claude models, requiring developers to run Claude Code via terminal or VS Code to use alternative providers.
How does configuring the settings.local.json file connect Claude Code to OpenRouter?
Users edit the environment block in settings.local.json by assigning the Anthropic base URL to the OpenRouter endpoint, inputting their OpenRouter API key as the auth token, and changing default model designations to the desired OpenRouter model name.
What causes the upstream idle timeout error when running large coding tasks through OpenRouter in Claude Code?
The error occurs when the model attempts to generate or inspect large volumes of text in a single request, exceeding the OpenRouter proxy timeout threshold before receiving a response.
How does the free stealth Ox Alpha model compare in execution speed and quality to Anthropic Claude models?
The model successfully creates functional code and detailed analytical reports, but it runs substantially slower than native Claude models, taking hours to finish builds that standard Claude completes in minutes.
Worth watching for
Developers and automation builders using Claude Code who want to reduce API costs by routing their coding assistant through free or alternative OpenRouter models.
- claude-code
- openrouter
- ai-agents
- vs-code
- llm
- prompt-engineering