DeepSeek Phone Harness: Remote-Control Your Desktop Agent from Your Phone

Open-source project DeepSeek Phone Harness — the full DeepSeek Harness experience on mobile: chat, model switching, tool-call approvals, terminal, file browsing — anywhere, over 4G/5G.

A familiar scene

It's 1 a.m. You're in bed. Your computer in the study is running a DeepSeek Harness task. Suddenly you want it to also do that refactor. So you get up, walk to the study, open the computer, type the instruction, watch it run, check the output, and walk back to bed — two minutes of walking just to send one sentence.

What if you could just pick up your phone, open a page, send the message, watch the streaming output, approve tool calls, even answer the agent's questions — and stay in bed?

That's exactly why I built DeepSeek Phone Harness: DeepSeek Harness in your pocket.

What it is

DeepSeek Phone Harness is an open-source mobile remote-control system. It lets your phone operate DeepSeek Harness on your computer from any network (4G/5G / WiFi / public internet):

Send tasks to your desktop agent and watch it work in real-time streaming
See every tool call: what it did, which file it touched, and the result
Approve risky operations: allow / reject
Answer agent questions so tasks never hang
Switch models, attach images, read the agent's thinking
Run terminal commands, browse files & code, review task history
The experience mirrors the DeepSeek Harness Web GUI

The UI follows the codex-relay design language (dark, card-based, gold whale branding), with the conversation experience aligned to DSH's native interface.

What it feels like

Chat, just like on your computer

Open the page and you're in a conversation: multi-turn sessions, Markdown rendering, streaming typewriter output. Switch models from the top (deepseek-v4-flash / v4-pro and more providers), attach images with the + button and send them to the agent.

Every step is visible

This is the part I'm most proud of — tool call cards:

[📄 read]                        ● done
C:/Users/Joey/Documents/.../config.json   ← path at a glance
▾ tap to expand:
  arguments  {"file_path": "...", "limit": 1}
  result     <path>... 1: {

Icons distinguish tool types (write / read / terminal / web / subagent…), status switches live (spinner while running → green done / red failed), the file path is shown directly, and tapping the card reveals the full arguments and result. The agent's thinking is also folded up for inspection, auto-expanding while streaming.

Approvals & questions — nothing hangs

DSH agents request authorization and ask questions. Both become cards on your phone:

Authorization needed: Allow once / Reject
Your answer needed: choose an option / type a custom answer / skip

Don't answer, it waits; answer, it continues — tasks never deadlock waiting for a human.

Terminal, files, tasks — the whole remote kit

Terminal: run commands on your computer from the phone, streaming output, full UTF-8, stoppable
Files / code: browse desktop directories, view file contents
Task history: review past tasks, expand output, re-run with one tap
Sessions: switch workspaces, create / rename / delete / search

What's interesting technically

Pure Node.js, zero dependencies: no npm packages at all — HTTP transport + DSH Web API RPC wrapper + events.mux WebSocket relay, all Node built-ins
Single-file frontend: relay.html, no framework, opens directly in the mobile browser
One WebSocket to rule them all: approvals, questions, and session events flow through a single events.mux connection; the frontend polls increments for the typewriter effect
Optional cloud: a Cloudflare Worker hosts the frontend + optional E2EE task channel; direct 4G/5G via Tailscale works too

Run it in three minutes

# On the computer (requires dsh web listening on 127.0.0.1:3080)
cp config.example.json config.json   # fill in your access token
node agent.mjs --mode=both