Source: raw/DeepSeek_Harness_Setup_-_A_Free_Claude_Code_You_Own_In_10_Minutes.md (auto-caption transcript)
Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=EDxVn1q8udE · Duration: 13:10 · Uploaded: 2026-08-28 · No sponsor named. Confidence is low: one creator’s hands-on with a developer preview; the repo, paper and star count were not checked.
DeepSeek Harness is DeepSeek’s coding-agent harness, which Sharbel installs and tests as “a free Claude Code you own”. The harness is free and the model is a plugin you choose, so you can run it on DeepSeek’s API, another provider, a local model or Anthropic’s API. Its distinguishing features in this video are that every component is a plugin you can switch off, that every run is fully traceable, and that a killed run resumes where it stopped.
Key Takeaways
- Model and harness are separate purchases. “The harness and the model were always two separate purchases and one of them just went to zero.” Claude Code is one harness where someone else made and closed the decisions; this one leaves them open.
- Install. One command from the site or repo, about five minutes; it opens a local web dashboard. It asks for a DeepSeek API key on first load, but you can skip it and add any provider in Settings, Models (he used his $10-a-month GLM subscription). Custom providers take an ID, display name, base URL and key, for example an Ollama endpoint.
- Everything is a plugin. Settings, Plugins lists every part, including the sidebar and commands, each with an on/off switch. A “creator mode” preset loads plugin-development tools. His test: “Add a plug-in that fires a Discord notification when a run finishes and puts the token count and what it cost me into the notification itself.” It built a notification bar, which needed two fixes (unreadable text colour, a broken counter).
- Traces. Each run’s trajectory shows the exact system prompt, the injected context word for word, which skills loaded and why, every tool call with payload and result, the thinking, and timings.
- Resume after a kill. He stopped a run mid-task and it continued with the same context, preferences and tool calls, because components declare what they need and the runtime keeps an undo for every change.
- It can drive Claude Code and Codex. Two packages (captions: “sub agent Claude Code” and “sub agent Codex”) spawn the real Claude Code and Codex as child processes and give the agent tools to hand them a task. Limits: one-shot calls that cannot be resumed; off by default (a preset must enable them); credential-shaped environment variables are stripped, so keys must be passed explicitly.
- Anthropic, OpenAI, Bedrock, Vertex and Codex are in the provider list, so you can run Opus inside the free harness, at API prices rather than on your Claude subscription.
- Status. A developer preview with only a release-candidate tag; the README warns of breaking changes in capitals.
- Free models have caps. Hosted free tiers are rate-limited (Google stopped publishing Gemini free-tier limits, he says); only local weights are truly unmetered.
Try It
- Install it, add a cheap provider (DeepSeek V4.1 Flash, GLM or a local Ollama model), and give it a week of work that does not need a frontier model. Measure how much of your month it covers.
- Open the trajectory of one run to see exactly what a harness sends to the model.
- Ask it to build one plugin you wish your current agent had.
Implementation
Tool/Service: DeepSeek Harness (DeepSeek), local web dashboard. Setup: Install command from the site or GitHub; providers in Settings, Models; plugins in Settings, Plugins. Cost: Harness free; you pay only the model provider. Integration notes: The Claude Code and Codex subagent packages need keys passed in explicitly and start a fresh process per call.
Open Questions
- He reports 190,000 GitHub stars (“the fastest growing repo on GitHub”), a formal paper and a framework called “Cordis”; none were checked.
- Licence and repository owner are not stated in the transcript.
- DeepSeek V4.1 Flash describes the harness as a free desktop app; this video shows a terminal install that opens a local web dashboard. They may be the same product in two forms
Related
- DeepSeek V4.1 Flash — the cheap model the harness is built around
- Chinese Models Decision Framework — DeepSeek’s free agent Search and Fetch tools through the harness
- Ollama + Claude Code — the other route: keep Claude Code, swap the model
- Codex Inside Claude Code — one harness driving another
- Harness Design for Long-Running Agents — what a harness has to provide
- Can a Mac Mini Replace an AI Subscription? (Sharbel) — the local-model side of the same cost question