DeepSeek changed the cost calculus for agent setups in late 2025. V3 handles general coordination tasks credibly at a fraction of frontier pricing, and R1 ships strong reasoning. For exploration agents, fact-checkers, classifiers, anything that produces text the user does not directly read, DeepSeek is often the right default.
app.yaml1brain:2 provider: deepseek3 model: deepseek-chat4 backend: openai_compat5 credential:6 ref: deepseek_main7 scope: per_user8 provider: deepseek9 config:10 api_key: "{{env.DEEPSEEK_API_KEY}}" # dev fallback11 base_url: "https://api.deepseek.com/v1"12 temperature: 0.2Register your DeepSeek key once in the encrypted credential vault, then reference it from any agent by name. The runtime resolves the secret at session start, never bakes it into the app bundle, and rotates without redeploys.
# one-time setup: register the credential
digitorn credentials create \
--provider deepseek \
--label deepseek_main \
--scope per_user \
-f api_key=...credential: block (shown above). The legacy {{env.X}} template still works as a development fallback.# 1. install the runtime
curl -sSL https://digitorn.ai/install | sh
# 2. register your DeepSeek credential in the vault
digitorn credentials create --provider deepseek --label deepseek_main -f api_key=...
# 3. install a starter agent and chat
digitorn install hub://digitorn/digitorn-code
digitorn chat digitorn-codeEngineering notes from the Digitorn team. No marketing, no launch announcements, no "10 prompts that will change your life". Just the things we write that we'd want to read.