Run DeepSeek's Super Agent locally for free with Ollama

Mervin Praison explains DeepSeek's research-prototype super agent (gateway, router, one worker per task) and sets it up for free on his own machine with the DeepSeek harness and a local Qwen 3.5 model in Ollama, no API key. He fixes a failing test in a demo repo.

Mervin Praison · YouTube · published Nov 8, 2026 · Software & DevOps

Run DeepSeek's Super Agent locally for free with Ollama — video by Mervin Praison▶ Watch on YouTube
Prompt
Fix the failing test in this repo. Read the code and the test, run the tests, find the bug, fix it and confirm the test passes.

Visual breakdown

DeepSeek's super agent is a research prototype that puts a router in front of many coding agents: every task gets its own worker, so you stop babysitting chats. Mervin sets it up free and fully local with Ollama and Qwen 3.5, and shows it fixing a failing test.

The problem: A person can only follow three to five AI sessions at once, so launching ten agents makes you the bottleneck.

How the super agent routes work

  1. Inbox messageEvery message or GitHub ticket you send lands in one inbox
  2. GatewayA plain-code guest list checks who sent it, before any AI reads it
    Dropped: Senders not on the list
  3. RouterReads each new message and decides where it goes, never does the work itself
  4. One worker per taskEach task gets its own agent with its own memory, working in its own folder
  5. Direct follow-upsFurther messages on the same task go straight to its worker, skipping the router
  6. Merge requestThe worker opens a merge request for you to review instead of you watching chats
InputsPlain codeAI agentResult

Super agent vs Claude Code agent teams

  1. Agent teams: one lead splits a big job into steps for several agents
  2. Super agent: delegates each whole task to a separate worker
  3. Best for many separate small tasks, such as bug reports and questions

By the numbers

1197tests passed on install
96plugins in the core
57tokens per second locally
63.3%tasks done, DeepSeek harness

Set it up free and local

Everything runs on your own computer through Ollama, so there is no paid API.

  1. Install Node.js 22.19 or later with PNPM, tmux, Python 3 and the GitHub CLI
  2. Install Ollama and run ollama pull qwen 3.5 9B (about 7 GB)
  3. Install the DeepSeek harness with npm -g, then git clone the super agent add-on, pnpm install and pnpm build
  4. Create a profile called Super Agent and edit config.patch.yml to point at Ollama and make Qwen 3.5 the default model
  5. Run the start script, open the login link from the log at 127.0.0.1:3082, then check Plugins and Models
  6. On the Super Agent settings card, enter your GitHub repository and the GitHub command the agent uses

Start in observation mode on a test project and let it earn your trust.

Caveats from the video

  • It is a research prototype, not an official DeepSeek product, and workers can run programs and read files in your home folder, so use a separate computer account.
  • DeepSeek's own test: 12.4 min vs 13.8 and 20.4, but about 23 million tokens vs 11 million and 6.1 million, so it costs more on paid models.
  • The speed test was preliminary and used a tool that just waited, so it is a promising design, not a proven win.
  • The agent ignores its own account, so real use needs a second GitHub account for it.
  • Mervin ran the manual task only; the full ticket-to-merge-request flow was DeepSeek's design, not his run.

Distilled from the video’s captions

How to try it

  1. Open DeepSeek.
  2. Copy the prompt and fill in the [placeholders].
  3. Watch Mervin Praison's video for the full walkthrough and setup.

Related use cases

Share