DeepSeek Harness setup and Claude Code comparison
Learn to configure DeepSeek Harness, connect model providers, and use its event log. The walkthrough starts a local web server on port 3080.
Player not loading? Watch on YouTube
DeepSeek Harness is an open source coding assistant runtime released under the MIT license. The walkthrough explains how its Cordis plugin framework lets users replace components such as the agent loop, storage and web interface. The speaker presents it as a configurable framework that requires setup and has rough edges.
Setup starts in a project directory with Node installed. The web interface runs at 127.0.0.1:3080. Users add an API key in Settings > Models, then choose a workspace before the message box accepts input. The tutorial also covers building from source with PNPM and connecting custom OpenAI-compatible endpoints or local models.
Four modes serve different tasks: Standard supplies the full tool set; Code orchestrates tool calls through programs; Minimal limits tools to a shell and file editor; Creator supports testing Cordis plugins in memory. Trajectory view records an append-only event stream that users can inspect, resume, fork or replay.
The speaker contrasts this self-hosted AI agent with Claude Code, citing a website-building test where DeepSeek finished sooner but used more tokens and produced a result the testers considered less polished. That example is not a general performance guarantee. The speaker also warns about unvetted plugins with shell access, recommends authentication before network exposure, and notes that version 0.1 expects compatibility-breaking changes.