Cursor, Copilot, a Claude subscription — the tool and the model come welded together. Price goes up, free tier dies, company changes its mind: you lose the whole thing and start over.
This build separates them. The agent lives on your machine and never changes. The model behind it is one line you can swap in ten seconds.
When you finish this, you own:
A coding agent running locally — reads your files, edits them, runs commands, keeps a plan. MIT licensed, built by DeepSeek, yours forever
Three separate fuel lines into it: free credits that refill daily · official DeepSeek at half price · open weights you can host yourself
The ability to swap fuel without rebuilding anything — same sessions, same plugins, same workflow
A 10-second test that tells you what any future “free AI credits” site really is
That last one is the point. Free API sites die — the one below has already announced it’s changing domains. When it does, you swap a URL and keep working. The build outlives its own links. 
The engine
1. Install it
node -v
npx @deepseek-ai/dsh web
Opens at http://127.0.0.1:3080.
DeepSeek Harness — official GitHub · MIT · 195k ★
Fuel line 1 — free credits that refill
2. Get a key
Register: Get Free AI Credits → copy your API key.
3. Plug it in
Settings → Models → Add a Custom Provider
Provider ID: hcnsec
Display name: HCNSEC
Base URL: https://api.hcnsec.cn/v1
API protocol: openai-completions
API key: YOUR_API_KEY
Fetch available models → pick → save.
You’re now coding on DeepSeek V4 Pro — 1M token context, the model that went GA on Aug 13. 
The 6 things that quietly break this
❶ “No models” / “no permission” — their own FAQ shouts it: create a NEW key and select your group on it. The default key ignores your entitlement. This is the #1 reason people say it doesn’t work.
❷ Free credits refill — if you click. The panel has daily check-in enabled. Signup bonus is once. Check-in is every day, forever, and nobody reopens the page. 
❸ Read the balance. 500,000 quota = $1. So 250,000 on screen is 50¢, not $250k.
❹ Model IDs went UPPERCASE and re-route to the official models. A saved config that suddenly 404s → re-fetch the list.
❺ GLM 5.3 may not be in your tier. Their announcements put it under SVIP; free users get a self-hosted qwen3.8. Don’t assume — fetch and read what your group returns.
❻ Screenshots get refused on custom providers. No checkbox exists. Edit $DSH_HOME/settings.yaml:
llm-pi-ai:
providers:
hcnsec:
models:
- id: YOUR-VISION-MODEL
input: [text, image]
Fuel line 2 — official, at half price
Add platform.deepseek.com as a second provider in the same harness. Both live side by side; pick per session.
Why bother when line 1 is free? Because peak is only 01:00–04:00 and 06:00–10:00 UTC, Mon–Fri. Every other hour — evenings, nights, all weekend — is 50% off:
V4 Pro, off-peak cached input $0.022 / 1M
input $0.66 / 1M
output $1.98 / 1M
Two cents per million cached tokens. Run the heavy jobs off-peak and this is closer to free than most “free” tiers.
Same trick on GLM-5.3: peak is 14:00–18:00 UTC+8 Mon–Fri only — weekends are half price all day.
Fuel line 3 — the one nobody can switch off
V4-Pro weights are open on Hugging Face. Not a demo, not a trial — the actual model. Rent a GPU by the hour, or wait for a community host. Same harness, new base URL.
And to judge every future site that appears:
Spot a relay in one command — no signup, works anywhere:
curl -s https://SITE/v1/models
If the error says "type":"new_api_error", it’s a New API panel — the open-source reseller software anyone can deploy in an hour. That’s exactly what the site above is. Not a scam by itself, but now you know what you signed up for instead of hoping.
Make it yours
| Trick | What it gives you |
|---|---|
dsh --profile headless "run the tests" |
One job, prints the answer, exits. Scriptable, no browser |
dsh web --port 8080 |
Move it off 3080 |
dsh web --no-open |
Running on a VPS over SSH — stops it fighting your browser |
--dump-config |
Shows the composed config before boot. Kills guesswork |
$DSH_HOME/.credentials.yaml |
Where keys really live. The UI never shows them again — back it up |
apiKeyEnv: MY_KEY |
Point at an env var instead of storing the key at all |
Provider ID is permanent — sessions and defaults reference it. Typo it and the only fix is delete + recreate.
Developer preview. The repo says breaking changes will land, and npx grabs the newest build every time. Pin a version once today’s setup works.
The plugin world — there’s no store, it’s a GitHub topic
github.com/topics/dsh-plugin — the whole ecosystem, ~11.6k repos
awesome-dsh-plugin — curated shortlist, start here
open-design — design plugin, ~91k ★
dsh-desktop — real desktop app instead of a tab
learn-harness-engineering — 0→1 if the plugin model confuses you
Discord · 
Site tutorial (CN)
Before you paste anything
Line 1 is a relay — a gateway forwarding your prompts to the real models. So:
Your code passes through someone else’s server. No private repos, no client work, no secrets.- Never put a real DeepSeek / OpenAI / Anthropic key into a relay.
That site has announced a domain change for licensing compliance. Configure fuel line 2 before you need it — that’s the entire reason it’s in this plan.
Everyone else is renting a coding tool and praying the price holds.
You’ll have the agent on disk, three ways to feed it, and a one-line test for the fourth.
The Core-Community used AI to refine the topic above.
!