How to Run a Free Model-Agnostic Coding Agent

TL;DR

Run the DeepSeek harness locally to create a free, customizable coding agent that can use DeepSeek, Claude, ChatGPT, or free OpenRouter models. Its MIT-licensed code lets you modify the interface, tools, memory, and connectors, while Claude Code or Codex may remain better choices when polished, client-facing output and dependable design quality matter most.

Transcript

Deepseek just became the fastest growing repo ever and that's because it's harness challenges Claude code. But the app is free, open source and runs any model that you want. And in this video I'll show you the new capabilities it unlocks, exactly how to use it so you can build and design with any model and any tools from $0. So if you haven't alrea... Read More

Key Insights

• A harness is the operational body around an AI model, while the model acts as the brain. It determines how the model uses tools, edits files, presents its work, manages sessions, and connects with external services.
• The DeepSeek harness is free, open source, MIT licensed, and designed to run on a user's machine. Its interface, plugins, chat tools, models, and underlying harness code can all be changed to fit a particular workflow.
• Model choice is independent of the harness, so users can select DeepSeek, Claude, ChatGPT, or free models through OpenRouter. A newly preferred model can be substituted without rebuilding the entire working environment.
• Local installation starts by copying the repository link from GitHub and asking a coding agent to download it and open it on localhost. After launch, the interface provides controls for selecting models, effort levels, workspaces, and operating modes.
• Custom tools can be built directly inside the harness. The demonstrated memory tool was created from a plain-language request and given functions to save, search, and forget information, with the harness showing its implementation activity and resource usage.
• Zapier can serve as an authentication and connection layer between the harness and personal services. The demonstrated integration exposed an email tool that allowed the agent to draft a Gmail message, while also supporting services such as Outlook.
• Standard mode supports ordinary coding-agent work such as editing files, while PTC mode exposes tools through a code-mode SDK for combined multi-step operations. Creative mode adds runtime inspection for building custom agent presets, and minimal mode offers another configurable option.
• Claude Code or Codex is presented as the safer choice when client-facing polish and design quality matter. The DeepSeek harness is positioned as a workhorse for high-volume, price-sensitive tasks where control, model flexibility, and lower cost are more important.
• More videos from this channel:
  • How to Use ChatGPT 5.6 Sol for Real Workflows
  • How to Give Claude Persistent Memory With Pinecone
  • How to use ChatGPT Voice 2.0 Across Five Levels
• More from this channel: Jack Roberts

Questions & Answers

Q: What is an AI coding harness?

An AI coding harness is the operational environment surrounding a language model. The model functions as the brain, while the harness acts as the body that lets it edit files, call tools, display progress, manage sessions, and connect with other services. This distinction matters because changing the harness can alter how the same model behaves and what practical tasks it can complete.

Q: How do you install the DeepSeek harness locally?

Open the project's website, follow its GitHub link, and copy the repository address from the code menu. The demonstrated method is to paste that address into a coding-agent chat and ask the agent to download the repository and open it on localhost. After installation, the local interface provides access to model selection, effort settings, workspaces, skills, tools, and configurable operating modes.

Q: Which AI models can the DeepSeek harness run?

The harness is described as model agnostic, meaning it can use different models within the same operating environment. The examples include DeepSeek, Claude, ChatGPT, and free models available through OpenRouter. This makes it possible to replace the selected model when preferences, performance requirements, or costs change, without abandoning the customized interface, connectors, tools, and workflows built around the harness.

Q: How can you add memory to the DeepSeek harness?

A memory capability can be added by asking the harness to build a memory tool for itself. In the demonstration, the resulting lightweight tool could save information, search stored information, or forget it. Because the harness is editable code, users can define how such memory behaves, inspect the implementation process, and review details such as completion time and token usage after the task finishes.

Q: How can the DeepSeek harness draft emails?

The harness can draft emails after it is connected to an external service through a tool integration. The demonstrated setup used Zapier as the authentication and connection layer, then exposed access to Gmail so the agent could create a draft describing Budapest. Zapier was selected because it could also connect with services such as Outlook and other personal platforms through one reusable integration layer.

Q: What operating modes are available in the DeepSeek harness?

Standard mode works as a regular coding agent and handles activities such as editing files. PTC mode includes the standard capabilities while exposing tools to a code-mode SDK, allowing the model to combine multiple operations in one TypeScript program. Creative mode adds runtime inspection for creating custom agent presets. Minimal mode is also available, although the demonstration mainly relies on standard mode.

Q: When should you use Claude Code or Codex instead?

Claude Code or Codex is recommended when output is client-facing, visual polish matters, or a website needs dependable design quality. The transcript argues that a model and its native harness contain more than raw intelligence, including design-related behavior that can improve results. Although the DeepSeek harness can be customized and systematized, Claude is presented as the more reliable choice for these polished deliverables.

Q: When is the DeepSeek harness the better choice?

The DeepSeek harness is a strong fit for high-volume work where price is important, for users who want to choose among multiple models, and for workflows requiring extensive customization. It runs locally, is free to use, carries an MIT license, and allows changes to its interface, plugins, tools, and underlying code. It can also connect to personal apps and support locally run models through Ollama.

Summary & Key Takeaways

• A model supplies the intelligence, while a harness supplies the operating environment through which that intelligence edits files, calls tools, manages workflows, and interacts with users. The DeepSeek harness offers a locally run, open-source alternative that can switch among DeepSeek, Claude, ChatGPT, and free models available through OpenRouter.

• Installation begins by opening the project's GitHub page, copying the repository link, and asking a coding agent to download the repository and launch it on localhost. Once running, users can choose models and effort levels, inspect activity and token usage, select workspaces, and modify the interface or underlying harness code.

• The harness can create custom capabilities, including a lightweight memory tool that saves, searches, or forgets information. It can also connect through Zapier to services such as Gmail or Outlook and draft emails. DeepSeek suits cost-sensitive, high-volume work, while Claude Code or Codex may deliver stronger polish for client-facing design projects.