@tailored-ai/provider-deepseek

DeepSeek model provider for Tailored AI agents — DeepSeek's V4 hybrid models, low-cost and strong.

DeepSeek speaks the OpenAI wire format, so chat, streaming, tool calling, and model discovery (listModels) come straight from core's OpenAIProvider. On top of that this plugin exposes DeepSeek's one non-standard knob — the per-request thinking toggle.

Install

tai plugin install @tailored-ai/provider-deepseek

Configure

plugins:
  - "@tailored-ai/provider-deepseek"

providers:
  deepseek:
    apiKey: "${DEEPSEEK_API_KEY}"          # https://platform.deepseek.com/api_keys
    defaultModel: "deepseek-v4-flash"
    thinking: false                        # non-thinking: fast, reliable tool calls

agent:
  defaultProvider: deepseek

Per-agent override — keep a local model for everyday work, reach for DeepSeek where it helps:

agents:
  researcher:
    provider: deepseek
    model: "deepseek-v4-pro"

| Field | Required | Notes | |---|---|---| | apiKey | yes | Use ${DEEPSEEK_API_KEY} to read from the environment. | | defaultModel | yes | deepseek-v4-flash or deepseek-v4-pro. | | thinking | no | false = non-thinking, true = thinking. Omit to use the model's native default (the V4 models think by default). | | baseUrl | no | Defaults to https://api.deepseek.com. Override for proxies. |

Models

| Model | What it is | Tool calling | |---|---|---| | deepseek-v4-flash | V4 hybrid, lower-latency tier. The right default for an agent. | Yes | | deepseek-v4-pro | V4 hybrid, most capable. For hard problems. | Yes |

Both are hybrids: a single model that either answers directly (non-thinking) or reasons first (thinking), selected per request by thinking.

• thinking: false — snappy, no reasoning tokens, reliable tool calls. This is the durable version of the old deepseek-chat behavior. The right default for tool-driven agent work.
• thinking: true — reasons before answering; stronger on hard problems, but slower and token-hungrier. Give it a generous maxTokens — a small budget gets eaten by the reasoning pass and the reply comes back empty with finish_reason: length.

The loop can still override the mode per call by passing extra: { thinking: { type: "enabled" } }.

A note on deepseek-chat / deepseek-reasoner

Those two model names are the non-thinking and thinking aliases of deepseek-v4-flash. DeepSeek deprecates them on 2026-07-24. Use deepseek-v4-flash with thinking instead — same model, no deprecation.

Not yet

In thinking mode the chain-of-thought (reasoning_content) isn't surfaced — only the final answer comes through. Streaming and tool calling work; the reasoning trace is dropped.

Development

pnpm --filter @tailored-ai/provider-deepseek run build
pnpm --filter @tailored-ai/provider-deepseek run test

License

MIT