Run DeepSeek Harness on a remote server

Start DeepSeek Harness#

On your computer, create a VM and connect. Pick any free name; if you've already clicked Deploy or have a VM, skip the first command and use its name:

ssh exe.dev new --name=my-dsh
ssh my-dsh.exe.xyz

On the VM, install Node.js 24 and DeepSeek Harness:

curl -fsSL https://deb.nodesource.com/setup_24.x | sudo -E bash -
sudo apt-get install -y nodejs
sudo npm install -g @deepseek-ai/dsh@0.2.1-rc.5

Point DeepSeek Harness at exe.dev's LLM gateway, which serves DeepSeek V4.1 Flash through Fireworks. The gateway needs no key, but DeepSeek Harness wants one set, so EXE_LLM_KEY is a placeholder:

mkdir -p ~/.dsh/profiles/web ~/project
echo 'EXE_LLM_KEY=not-a-secret' >> ~/.dsh/.env
cat > ~/.dsh/profiles/web/cordis.patch.yml <<'EOF'
# DeepSeek V4.1 Flash through exe.dev's LLM integration: no API key on the VM.
- id: llm-pi-ai
  config:
    providers:
      exe-llm:
        displayName: exe.dev
        api: openai-completions
        baseURL: https://llm.int.exe.xyz/fireworks/inference/v1
        apiKeyEnv: EXE_LLM_KEY
        models:
          - id: accounts/fireworks/models/deepseek-v4p1-flash-1017
            name: DeepSeek V4.1 Flash 1017
            contextWindow: 1048576
            input: [ text, image ]
            compat:
              thinkingFormat: deepseek
            reasoningEfforts:
              off:
              high: high
              max: max
          - id: accounts/fireworks/models/deepseek-v4p1-flash
            name: DeepSeek V4.1 Flash
            contextWindow: 1048576
            input: [ text, image ]
            compat:
              thinkingFormat: deepseek
            reasoningEfforts:
              off:
              high: high
              max: max
- id: agent-default-model
  config:
    provider: exe-llm
    model: accounts/fireworks/models/deepseek-v4p1-flash-1017
EOF

Check that a model answers. dsh headless uses its own profile, so pass the web profile's settings:

cd ~/project
dsh headless --patch ~/.dsh/profiles/web/cordis.patch.yml "Reply with exactly: OK"

Open it in your browser#

Run dsh web as a service, so it keeps going when you disconnect and starts again after a reboot. It stays on its default address, 127.0.0.1:3080, and exe.dev's proxy brings it to https://my-dsh.exe.xyz:3080/, behind your exe.dev login. --trusted-host lets that address through DeepSeek Harness's own host check:

sudo tee /etc/systemd/system/dsh-web.service >/dev/null <<'EOF'
[Unit]
Description=DeepSeek Harness Web UI
After=network-online.target
Wants=network-online.target

[Service]
User=exedev
WorkingDirectory=/home/exedev/project
ExecStart=/bin/bash -lc 'exec dsh web --no-open --trusted-host "$(hostname).exe.xyz" --public-url "https://$(hostname).exe.xyz:3080/"'
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target
EOF
sudo systemctl daemon-reload
sudo systemctl enable --now dsh-web

Print the sign-in link:

journalctl -u dsh-web -o cat | grep 'dsh web:' | tail -1

Open it. The token in the link signs your browser in for 30 days, and works until the service restarts; after that, print the new one the same way. Accept the preview notice, open the project folder as your workspace, and give it a task, such as "Write index.html for a one-page website for a bakery that only sells sourdough."

Close the tab whenever you like; running tasks keep going on the VM.

Use your own DeepSeek key#

Optional. DeepSeek V4-Pro and DeepSeek's own API need a DeepSeek API key. Keep it off the VM with an HTTP proxy integration that adds it to every request. On your computer:

ssh exe.dev integrations add http-proxy --name deepseek --target https://api.deepseek.com --header x-api-key:YOUR_DEEPSEEK_KEY --attach vm:my-dsh

On the VM, send DeepSeek Harness's DeepSeek provider through it, with a placeholder key, and restart:

echo 'DEEPSEEK_API_KEY=not-a-secret' >> ~/.dsh/.env
cat >> ~/.dsh/profiles/web/cordis.patch.yml <<'EOF'
- id: llm-deepseek
  config:
    baseURL: https://deepseek.int.exe.xyz/anthropic
EOF
sudo systemctl restart dsh-web

DeepSeek-V4.1-Flash and DeepSeek-V4-Pro then show up under DeepSeek in the model picker. We checked that DeepSeek Harness sends its requests and key header this way, but haven't run it with a real DeepSeek key.

Add the superagent#

Optional, and early. On Oct 31, DeepSeek published dsh-superagent, a service that watches GitHub notifications for a dedicated agent account and answers issues and mentions with comments and pull requests. Its README calls it a research prototype, not an official DeepSeek product. It needs a separate GitHub user account (not an app) on github.com.

It runs as a second DeepSeek Harness profile on port 3082. On the VM, install pnpm and build it:

sudo npm install -g pnpm@11
git clone https://github.com/deepseek-ai/dsh-superagent-prototype.git ~/dsh-superagent
cd ~/dsh-superagent
pnpm install && pnpm build
dsh superagent --from-default-profile web --dump-config > /dev/null
dsh plugin --profile superagent add "link:$PWD"

Give the new profile the same model settings, and the port and state folder from its deployment guide:

mkdir -p ~/.dsh/state/superagent
{ cat ~/.dsh/profiles/web/cordis.patch.yml; cat <<EOF
- id: webserver
  config:
    host: 127.0.0.1
    port: 3082
- id: dsh-superagent
  config:
    stateDir: $HOME/.dsh/state/superagent
EOF
} > ~/.dsh/profiles/superagent/cordis.patch.yml

Its scripts/dsh-superagent helper finds the service by process name, and Node.js 24 names it MainThread, so status reports a running service as down. Until that's fixed upstream, patch your copy:

sed -i "s/('node', 'dsh')/('node', 'dsh', 'MainThread')/" scripts/dsh-superagent

The rest is accounts and settings, which its SETUP.md has an agent walk you through one step at a time. Ask DeepSeek Harness (or Shelley) on the VM: "Read SETUP.md in ~/dsh-superagent and set up dsh-superagent with me." It has you create the agent's GitHub account and sign it in with gh, starts the service in tmux, and sets up an hourly check that restarts it.

Its settings panel only accepts requests to localhost, so open it through an SSH tunnel. On your computer:

ssh -L 3082:localhost:3082 my-dsh.exe.xyz

In that session, print its sign-in link and open it on your computer:

grep -h 'dsh web:' ~/dsh-superagent/.artifacts/*.log | tail -1

We built and started the service on a VM and it waited for setup as expected; we didn't create a GitHub account for it, so we haven't seen it answer an issue.