Run DeepSeek Super Agent Locally for FREE (Harness + Ollama)
DeepSeek's Super Agent, explained simply, then set up free and 100% local: no API key, nothing to pay. A router hands each job to its own AI worker, the workers run at the same time, and a follow-up goes straight to the worker that owns it. I installed DeepSeek's prototype, checked their numbers (and the catch), showed how DeepSeek Harness is built from plugins, and ran it on a free local model, qwen3.5:9b in Ollama. With one GitHub account Super Agent's GitHub flow does not run: it needs a second account for the agent. Timestamps 0:00 DeepSeek's Super Agent, free on your computer 0:32 What you'll learn 0:56 The problem: you can only follow 3 to 5 AI chats 1:33 The idea: a restaurant (router, workers, gate) 2:47 DeepSeek's prototype, installed 3:50 What DeepSeek's numbers really say 6:16 Super Agent vs Claude Code agent teams 6:47 How DeepSeek Harness works: everything is a plugin 8:08 Set it up free and local (Ollama + qwen3.5) 13:03 DeepSeek Harness fixes a bug, and the one-account test 14:36 When to use it, and staying safe 15:07 Recap Setup (as in the video; tested from a fresh folder, 7 Nov 2026) 1. Free tools: Node.js 22.19+, pnpm, git, tmux, Python 3, GitHub CLI (gh): node -v; pnpm -v; git --version; tmux -V; python3 --version; gh --version 2. Ollama from ollama.com, then: ollama pull qwen3.5:9b 3. DeepSeek Harness: npm install -g @deepseek-ai/dsh@0.2.1-rc.3 4. The Super Agent add-on: git clone --depth 1 https://github.com/deepseek-ai/dsh-superagent-prototype.git cd dsh-superagent-prototype pnpm install pnpm build 5. The profile, then link the add-on into it. Run the first line once only: run again, it resets the profile (your settings and the add-on link). The video hides its output: dsh superagent --from-default-profile web --dump-config dsh plugin --profile superagent add "link:$PWD" 6. Edit the profile's settings file: nano ~/.dsh/profiles/superagent/cordis.patch.yml Delete the last line, [], paste this (use your own home folder), save with Ctrl+O, Enter, close with Ctrl+X: # The port and the folder Super Agent keeps its data in. - id: webserver config: host: 127.0.0.1 port: 3082 - id: dsh-superagent config: stateDir: /Users/praison/.dsh/state/superagent # Use a local Ollama model (free, no API key). - id: llm-pi-ai config: providers: ollama: displayName: Ollama (local) api: openai-completions baseURL: http://127.0.0.1:11434/v1 apiKeyEnv: OLLAMA_API_KEY # any value works compat: supportsDeveloperRole: false maxTokensField: max_tokens models: - id: "qwen3.5:9b" contextWindow: 65536 - id: agent-default-model config: {provider: ollama, model: "qwen3.5:9b"} 7. export OLLAMA_API_KEY=ollama 8. scripts/dsh-superagent start 9. Print the login link and paste it into your browser's address bar (127.0.0.1:3082 is your own computer): grep -ho 'http://127.0.0.1:3082/?token=[^ ]*' .artifacts/*.log | tail -1 10. Continue. More, Settings, Models: Ollama (local) has a green dot. Plugins, Superagent: 5 parts, code source link: the folder you cloned. Superagent, Settings, Sources: your repository and gh, then Save. A real run needs a second GitHub account for the agent (docs/006-deployment.md). Workers run with full access as your user. How DeepSeek Harness works - Everything is a plugin: model, shell, web search, tools, even the agent loop. The core has 96. - A profile stacks bundles (core, web app, Super Agent); your cordis.patch.yml is on top, so it wins. - Super Agent is one bundle of 5 parts; DeepSeek did not change the harness. My tests (7 Nov 2026) - Mac, 32 GB. qwen3.5:9b fixed the demo bug in 3 of 3 runs (47 s to 3 min 9 s), 57 tokens/s, no API key. - DeepSeek's numbers (section 7.1, preliminary, simulated work): 12.4 min a task (Super Agent) vs 13.8 (told to delegate) vs 20.4 (plain prompt); tokens 23.2M vs 10.8M vs 6.1M. - FrontierHarness Eval (not DeepSeek's chart; Kimi K3, 30 tasks): DeepSeek Harness 63.3% at $3.28 a task, Claude Code 63.3% at $18.34, Codex 66.7% at $3.47. It tested the harness, not Super Agent. Sources - DeepSeek, Exploring Super-agent Design: https://x.com/DeepSeekHarness/status/2116857257245987236 - Super Agent prototype: https://github.com/deepseek-ai/dsh-superagent-prototype - DeepSeek Harness: https://github.com/deepseek-ai/deepseek-harness - FrontierHarness Eval: https://frontierharness.org - Claude Code agent teams: https://code.claude.com/docs/en/agent-teams - My demo repository: https://github.com/MervinPraison/superagent-demo What's the first job you'd hand your own worker? Tell me in the comments. #DeepSeek #SuperAgent #Ollama
