• Open-sourced 13 Aug 2026, now in public preview — github.com/deepseek-ai/deepseek-harness
• Five profiles: web, tui, headless, sdk, acp (the last speaks to editors like Zed)
• Desktop app for macOS + Windows (v0.2 preview); Linux lives in the CLI
• No first-party VS Code / JetBrains extension. Marketplace hits are community-built — quality unverified

# the whole install
node -v        # must be ^22.19 or >=24 — it HANGS on 20
npx --yes @deepseek-ai/dsh --version
npx @deepseek-ai/dsh web      # → 127.0.0.1:3080

# one-shot, no UI
npx @deepseek-ai/dsh --profile headless \
  "summarise this repo and list its packages"

# see the plugin tree (the party trick)
npx @deepseek-ai/dsh web --dump-config | head -40

The silent Node-20 hang is the single most common "dsh is broken" report. Check node -v first, always.

What it's genuinely good at

• Cost per solved task. Terminal-Bench 4.0: $0.54 per test vs $13.20 (Claude Opus 5.5) and $9.58 (GPT-6 Astra) — an 18–24× gap
• 1M context, 384K output — you can paste a real repo in
• Index 39, #10 of 121 on Artificial Analysis, ~181 tok/s
• MIT weights — you can self-host later if you ever want to

What will annoy you

• Verbosity. 250M output tokens on that index — the highest of 47 models measured (median 140M). Output is 4× input price, so verbosity is your bill
• Instruction drift on long files — the 1M window is real, but quality fades well before you fill it. Keep files small; give it fresh context
• Anthropic-layer gaps: cache_control ignored, document blocks unsupported, MCP tool blocks unsupported — Claude-Code + MCP workflows degrade
• Vendor benchmark caution: DeepSeek's own 42.3% uses the Claude Code harness; the independent 31.8% uses Mini-SWE-agent. Both true, different claims

# 1 node must be ^22.19 or >=24 (20 hangs silently)
node -v
# 2 key from platform.deepseek.com/api_keys
export DEEPSEEK_API_KEY=sk-...
# 3 smoke test without installing anything
npx --yes @deepseek-ai/dsh --version
# 4 boot the web UI
npx @deepseek-ai/dsh web        # 127.0.0.1:3080
# 5 raw API sanity check
curl -s https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-flash","messages":[{"role":"user","content":"hi"}]}'

# 6 check balance — catches 402 before it bites
curl -s https://api.deepseek.com/user/balance \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY"

# 7 same model inside Claude Code
npm i -g @anthropic-ai/claude-code
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=$DEEPSEEK_API_KEY
export ANTHROPIC_MODEL="deepseek-flash[1m]"
claude

# 8 or inside OpenCode (cleaner, no env vars)
opencode           # then: /connect → deepseek → key

• Credits are token-weighted: (in×w + cached×w + out×w)/10,000
• GLM-5.3 weights: 6.9 input · 1.7 cached · 24 output — output dominates
• Two caps at once: a 5-hour bucket and a weekly bucket
• Weekends cost 50% all day; peak is Mon–Fri 14:00–18:00 SGT

• An "ADE" (agentic development environment) — Electron desktop app: .dmg / .exe / AppImage / deb / rpm
• The app is free. You bring a Coding Plan or any Anthropic/OpenAI-compatible key
• Built-in: file tree, terminal, git commit, live browser preview, review-diff flow, long-running "Goals", mobile remote, Telegram bot control
• No plan? 5-day trial at 8M tokens/day (3M GLM-5.3 + 5M Flash), then bring a key

# install
macOS    drag ZCode.app → /Applications
         # Gatekeeper: xattr -dr com.apple.quarantine ZCode.app
Windows  run the .exe installer
Linux    chmod +x ZCode-*.AppImage   # needs libfuse2
         # or the .deb / .rpm

Fine for learning the workflow. Less ideal as your permanent setup, because model routing is GLM-first and the app is desktop-only.

# install (pick one)
curl -fsSL https://opencode.ai/install | bash
npm i -g opencode-ai
brew install anomalyco/tap/opencode
choco install opencode       # windows (WSL recommended)

# point it at a provider
opencode
/connect          # → deepseek | z.ai coding plan | anthropic | openrouter
/models           # pick; switch per-session

Keys land in ~/.local/share/opencode/auth.json. 75+ providers via models.dev — the same harness works whether you're on a ₹500/month token budget or a corporate key.

// ~/.config/opencode/opencode.json — any base URL
{
  "$schema": "https://opencode.ai/config.json",
  "model": "deepseek/deepseek-flash",
  "provider": {
    "zai": {
      "options": { "baseURL": "https://api.z.ai/api/anthropic" }
    }
  }
}

• Build / Plan agents — Tab switches; Plan won't touch files
• Subagents for search/explore, @-mention files
• LSP diagnostics ("lsp": true) — real type errors, not vibes
• /init writes AGENTS.md; /share gives a public link
• VS Code / Cursor extension, Cmd+Esc

• Week 1 — OpenCode + a ₹500 DeepSeek top-up. Learn the loop: /init, Plan agent, review every diff before accepting. Cost: ₹500, refundable as unused credit.
• Week 2 — Add AGENTS.md to each project (OpenCode's /init writes it). This is the single biggest quality jump, and it's free.
• Week 3 — Try ZCode's free 5-day trial only to see whether you prefer GUI over TUI. Uninstall if not.
• Only then — if your monthly token spend crosses ~₹1,700 and you want it fixed, buy GLM Lite. Not before.