Course Overview
This course teaches how to run DeepSeek Harness locally using free Gemini and OpenRouter AI models to build custom tools and automate coding workflows without any paid API keys or GPUs. It covers configuring model providers, controlling agent permissions, reading session trajectories, building reusable plugins, and running agents headlessly for automation.
Students will complete projects including a repository explorer, a controlled bug fix, a custom project-audit tool, and a capstone repository maintenance agent that produces verified test evidence of its work.
Key Takeaways
- Install and run DeepSeek Harness on Windows, macOS, or Linux without GPU or paid API keys.
- Configure free Gemini and OpenRouter model providers and switch between them seamlessly.
- Control agent access to files and terminal with permission modes and approval workflows.
- Read and interpret session trajectories to verify agent claims with real test evidence.
- Build persistent project instructions, reusable skills, custom commands, and plugins.
- Run the harness headlessly and automate workflows programmatically using Python.
- Understand when multiple agents add value versus duplicating work.
- Complete a capstone project delivering a fully verified repository maintenance agent.
Prerequisites
- Ordinary Windows, macOS, or Linux computer (no GPU required).
- Node.js version 22.19 or 24 and up (installation instructions included).
- Comfortable using a terminal and running commands.
- Ability to read small JavaScript files to recognize bugs.
- Google account for Gemini API and OpenRouter account (both free to create).
- No prior AI or agent experience required.
Target Learners
- Developers facing paywalls on AI coding tools seeking free alternatives.
- Junior and mid-level engineers wanting to understand coding agents deeply.
- Students and self-taught developers needing zero-cost real agent experience.
- Developers in regions without practical paid API access.
- Platform and DevOps engineers evaluating agent harnesses.
- Technical leads defining AI agent policies.
- Anyone wanting to verify agent claims with real evidence.
Welcome and course outcome 3:15
Model versus agent versus harness 2:51
How the zero-cost architecture works 2:55
Developer , telemetry and safety 3:30
Development-environment requirements 2:44
Install Node.js and verify the environment 2:11
DeepSeek Harness with NPX 2:04
Create a safe practice workspace 2:06
Navigate the DeepSeek Harness Web UI 1:44
Diagnose installation and startup failures 2:37
Harness architecture and safety
Assignment: Prove your environment and bound your workspace before any model touches it
Choosing a provider that can run an agent 3:00
Create a Google AI Studio API key 3:00
Configure Gemini in DeepSeek Harness 2:03
Run your first tool-enabled session 2:53
Gemini free-tier limits and privacy 2:24
Troubleshoot Gemini connections 2:38
Assignment: Configure a free provider and inspect a repo without changing it
Why add a second model provider 2:22
Create and secure an OpenRouter API key 2:34
Configure OpenRouter in DeepSeek Harness 1:49
Select free coding-agent models 2:22
Why not rely on automatic free routing 2:21
Run the same task with Gemini and OpenRouter 2:37
Provider configuration
Assignment: Run one task through two free models and compare them
Give the agent a precise coding objective 2:37
Repository exploration with read-only tools 2:02
Use Plan Mode before implementation 2:14
Approve controlled file modifications 2:22
Run tests and collect evidence 1:42
Continue after a failed test 2:38
Inspect the complete trajectory 2:33
Tool calls and trajectories
Assignment: Fix a reported defect and prove the fix with a regression test
Persistent project instructions 3:00
Build a reusable repository-inspection skill 2:43
Create a custom command 2:35
Understand the plugin architecture 2:30
Build a simple custom tool 2:21
Package and reuse the plugin 2:11
Plugins and reusable workflows
Assignment: Build a project-audit tool and prove it is reusable
When to use one agent or several 2:29
Delegate repository analysis to a child agent 2:21
Build an implementation-and-review workflow 2:15
Run DeepSeek Harness in headless mode 2:24
Use the Python SDK 2:31
Design a repeatable maintenance workflow 2:30
Automation and multi-agent design
Assignment: Run an implementation-and-review workflow against an unreported defect
Capstone requirements and acceptance criteria 2:11
Phase 1 — Inspect and diagnose 1:59
Phase 2 — Implement under approval controls 2:27
Phase 3 — Test and recover from failure 2:11
Phase 4 — Produce an auditable completion report 2:47
Capstone: Repository Maintenance Agent, issue intake to verified report
Cloud inference versus fully local inference 2:30
Install and test Ollama 1:40
Connect Ollama to DeepSeek Harness 2:48
Test a small local model 2:52
Understand local-model limitations 2:51
Quiz — Running locally and judging the result
Assignment: Decide whether a local model is good enough, on evidence
Choose the right model route 1:54
Production-readiness checklist 1:44
Final challenge and next steps 2:24
Assignment: The final challenge — extend the work into something nobody briefed you on
Access Files No pi required.
Free Course files 1 file package
- Full Pack
Discussions are closed
Comments are currently disabled for this course.
