Your project.
Always monitored.
Carefully automated.
Build a reproducible DeepSeek Super Agent setup on Windows 11, keep its Linux environment on an external SSD, and connect an OpenRouter model. Choose an AI-assisted or manual installation path.
Community installation guide • Not affiliated with DeepSeek • Last reviewed: November 4, 2026
Event
Policy
Ticket
Understand the setup
Super Agent is a research prototype that runs as a DeepSeek Harness bundle, not a standalone model. Its worker presets have broad OS access; don't point them at your personal files.
Dedicated external SSD
Place Ubuntu's virtual disk and the Linux-side tools, Harness state and workspaces on the SSD. Some WSL host components remain on Windows.
Separate GitHub identity
Use a disposable repository and a dedicated bot user. Do not give the worker your primary GitHub credentials.
Human approval
Leave auto-merge disabled and proactive acceptance in observe mode. Review changes and logs yourself.
Pick your installation path
Both paths end in the same Harness + Super Agent installation. Commands are provided for a Windows 11 host and an Ubuntu WSL2 environment on an external SSD.
Step zero — identify the external SSD
Open My Computer → right-click the SSD → Properties. Confirm its drive letter, free space, and filesystem. Never format the drive. The examples use E: only as a placeholder.
Choose a folder name or a relative subfolder path. The install commands and Codex prompt update automatically.
E:\TunnelLab-AI\Ubuntuwsl --version
wsl --list --verbose
Get-VolumeAI-assisted: let Codex handle the setup
Run Codex Desktop in a local environment that can actually access Windows and WSL. Paste this prompt into Codex. It performs discovery first and asks permission before any modifications.
You are my local Windows 11 installation agent. Set up DeepSeek Super Agent on my external SSD from preflight through an inactive local panel using the TunnelLab single-prompt workflow. Read and follow this complete, version-pinned operator guide first:
https://github.com/pouramin/dshsuperagent/blob/main/installer/CODEX_PROMPT.md
Official upstream remains https://github.com/deepseek-ai/dsh-superagent-prototype — TunnelLab is independent.
My external SSD drive is E: (verify this yourself; it may be different).
Goal: store the Ubuntu WSL virtual disk, Harness packages, DSH_HOME, source checkout, caches, state, worktrees and logs on that SSD, not the internal C: drive whenever supported.
One conversation, no repeated prompts for normal stages. First verify the exact disk and present the complete safety plan for approval. Obtain separate authorization for administrator changes or any potentially forced reboot. NEVER reboot Windows, interrupt OBS, format disks, fall back to C:, overwrite an existing profile, expose credentials, or enable workers / auto-merge. Resume safely after a manual reboot. Stop at the first authenticated panel and hand over GitHub/model setup to me. If you cannot read the full guide, stop instead of improvising.Codex cannot approve Windows UAC prompts, create accounts or supply private API keys on your behalf. Those are manual checkpoints.
Manual setup: exact commands and checkpoints
Run PowerShell commands in Windows and bash commands inside Ubuntu. If the named distribution already exists, stop and inspect it instead of reinstalling or overwriting it.
powershell -NoProfile -ExecutionPolicy Bypass -File .\installer\install.ps1 -DriveLetter E -Folder TunnelLab-AIDownload and inspect installer files ↗
Enable WSL without a default distribution
wsl --install --no-distribution
wsl --update
wsl --list --onlineRestart Windows if prompted. Confirm that Ubuntu-24.04 appears in the online list before using that distribution name.
Install Ubuntu onto the SSD
wsl --install -d Ubuntu-24.04 --location "E:\TunnelLab-AI\Ubuntu" --no-launch
wsl --list --verboseOn WSL builds that do not recognize --location, stop and update WSL; don't fall back to a default C: installation. For a dedicated restricted Linux user, follow the installer script or the security notes below before starting any agents.
Install basic Linux tools
sudo apt update
sudo apt install -y git python3 python3-venv tmux curl ca-certificates build-essential gh
# Install Node.js 24 using the maintained NVM project first:
# https://github.com/nvm-sh/nvm#installing-and-updating
# In a new Ubuntu shell, after NVM is installed:
nvm install 24
nvm use 24
npm install -g pnpm@11.7.0
node -v
pnpm -v
git --version
tmux -V
python3 --version
gh --versionThe final commands check versions. Install Node.js 22.19+ (or 24+) and pnpm using their official Linux instructions, not an outdated Ubuntu package. Git must be 2.32+ and gh 2.48+.
Install the pinned Harness version
npm install -g @deepseek-ai/dsh@0.2.1-rc.3
dsh --versionThe version is pinned by the upstream Super Agent deployment guide. Confirm compatibility before upgrading.
Prepare the project and dedicated profile
mkdir -p ~/src
cd ~/src
git clone https://github.com/deepseek-ai/dsh-superagent-prototype.git dsh-superagent
cd dsh-superagent
git checkout 8c32a9dc5d72c033816c6a8110b37f40cf94fba4
pnpm install --frozen-lockfile
pnpm typecheck
pnpm build
export DSH_HOME="$HOME/.local/share/dsh"
mkdir -p "$DSH_HOME"
# Only for a NEW profile — stop if superagent already exists:
test ! -e "$DSH_HOME/profiles/superagent" || { echo "Profile already exists"; exit 1; }
dsh superagent --from-default-profile web --dump-config > /dev/null
dsh plugin --profile superagent add "link:$PWD"Run profile creation ONLY once on a new installation. On an existing profile, the creation command can overwrite settings; use the official Upgrade guide instead.
Configure the machine profile
Copy examples/profile.cordis.patch.yml to $DSH_HOME/profiles/superagent/cordis.patch.yml. Replace the stateDir placeholder with an absolute path (for example $DSH_HOME/state/superagent); use localhost port 3082. Set the dedicated GitHub CLI later in Settings → Sources, not in this YAML. Preserve existing configurations; do not overwrite a previously configured profile.
Official profile example ↗Configure a separate GitHub bot identity
mkdir -p ~/.local/bin
cat > ~/.local/bin/gh-agent <<'EOF'
#!/bin/sh
GH_CONFIG_DIR="$HOME/.config/gh-agent" exec gh "$@"
EOF
chmod +x ~/.local/bin/gh-agent
gh-agent auth login
gh-agent api user --jq .loginCreate the dedicated GitHub user and accept the test repository invitation manually. Only sign in as the bot for gh-agent. Don't reuse your main GitHub CLI credentials.
Start and configure Super Agent
scripts/dsh-superagent start
scripts/dsh-superagent statusOpen the panel URL shown in the start output; select Superagent in the sidebar. Set Sources, Policy, Execution and Prompts in that order. If status reports setup pending (exit 3), finish the missing settings rather than restarting repeatedly. If Node 24 reports MainThread and the process-detection script times out, verify the real process identity before any reviewed fix.
Prepare a separate test checkout and a safe first test
Initialize a disposable GitHub repository with README, invite a dedicated GitHub bot, clone with that bot's isolated gh-agent identity, and configure Settings → Execution → Prepare command from upstream Deployment. Set Eval and Execution to the working model (for example openrouter/free), select the same model in Open main Session, pause Executor and leave auto-merge off. Receive-mode Observe does not prevent explicitly addressed requests from becoming tickets; confirm safeguards before retrying.
Connect OpenRouter
Harness supports a custom OpenAI Chat Completions-compatible provider. Free OpenRouter models may work, but tool use, context limits and reliability vary; compatibility must be tested rather than assumed.
Settings → Models → Custom model API
Provider ID: openrouter
Base URL: https://openrouter.ai/api/v1
API protocol: openai-completions
API Key: enter privately in Harness
Model ID: openrouter/free (automatic free-model routing; verify tool calling)Select Fetch available models, check tool-calling support and test a small task. A free API tier does not guarantee unlimited requests or uninterrupted operation.
Costs and limits
The prototype's source code is MIT-licensed. Using a free model may cost $0 in model tokens within the provider's applicable limits. Storage, networking, paid models, retries and hosted compute can still cost money.
Verify before trusting the agent
Use a disposable GitHub repository with a trivial reproducible bug. Mention the dedicated bot in a NEW issue comment, then watch the Ticket and review every proposed change.
Sources & documentation
Documentation can change. These are the primary sources used for this installation guide; verify version-specific commands before applying upgrades.