n8n-nodes-deepseek-ai-chat

DeepSeek V4-native Chat Model community node for self-hosted n8n.

Connect DeepSeek AI Chat Model to AI Agent, Basic LLM Chain, or any node that accepts an ai_languageModel connection.

| | | |---|---| | npm package | n8n-nodes-deepseek-ai-chat | | Version | 0.1.3 | | Node type | n8n-nodes-deepseek-ai-chat.lmChatDeepSeekAi | | Display name | DeepSeek AI Chat Model | | Credential | DeepSeek AI Chat API (deepSeekAiChatApi) | | License | MIT | | Author | Danielnke |

Status

| Item | Status | |---|---| | Phase 1 Chat Model sub-node | Implemented | | Credentials + model list | Implemented | | Thinking mode + reasoning effort | Implemented | | Host LangChain resolution (dual-package fix) | Implemented (0.1.1+) | | Sub-node tracing (output panel + running glow) | Implemented (0.1.2+) | | Cost / latency / cache meta on model panel | Implemented (0.1.3+) | | Build / TypeScript | Passes (npm run build, npm run lint) | | Git repository | Initialized (main) | | npm publish | Not published yet (install from source) | | Phase 2 action node (message / classify / FIM) | Not started |

Why this package?

n8n’s built-in DeepSeek Chat Model works but lags DeepSeek V4 product needs:

• Models: deepseek-v4-flash, deepseek-v4-pro
• Explicit thinking toggle (API defaults thinking on; we default UI to disabled for safer agents/tools)
• Reasoning effort (high / max)
• Visible base URL (including /beta when needed)
• No deprecated frequency/presence penalty UI
• Sub-node UX parity with built-ins: glow while calling + raw I/O in the model panel

Legacy aliases deepseek-chat / deepseek-reasoner are deprecated by DeepSeek (see API docs).

Features

• OpenAI-compatible chat via LangChain ChatOpenAI + DeepSeek base URL
• supplyData Chat Model sub-node (ai_languageModel)
• Static V4 model options + dynamic load from GET /models
• Thinking mode always sent as explicit thinking.type (DeepSeek V4 default is enabled if omitted)
• Options: temperature, top P, max tokens, JSON mode, timeout, max retries, stop, user id
• Credential test: GET /models
• Host module resolution for @langchain/openai (avoids dual-package instanceof breaks with Basic LLM Chain)
• LLM tracing callbacks (host N8nLlmTracing when available, else fallback):
  • addInputData → running glow
  • addOutputData → panel output (messages, generations, token usage)

Prerequisites

• Self-hosted n8n with LangChain AI nodes (roughly 1.77+; tested around 2.x)
• DeepSeek API key: platform.deepseek.com/api_keys
• Node.js for building (package uses TypeScript 5.9 / modern n8n peers)

n8n Cloud / verified community: This package has runtime LangChain deps and is intended for self-hosted installs. Cloud verification generally disallows external runtime dependencies.

Installation

From source (current)

cd n8n-nodes-deepseek-ai-chat
npm install
npm run build

Load the package root (folder with package.json + dist/) into n8n, then restart n8n.

Common patterns:

# Custom extensions path (process / Docker)
N8N_CUSTOM_EXTENSIONS=/path/to/n8n-nodes-deepseek-ai-chat

Or install/copy per n8n community / private nodes docs.

After restart, search the palette for DeepSeek AI Chat Model.

From npm (when published)

Settings → Community Nodes → Install
Package: n8n-nodes-deepseek-ai-chat

Not on the registry until you run npm publish.

Credentials

Create DeepSeek AI Chat API:

| Field | Default | Notes | |---|---|---| | API Key | — | From DeepSeek platform (password field) | | Base URL | https://api.deepseek.com | Use https://api.deepseek.com/beta only for beta features |

Credential test calls GET {baseURL}/models.

Usage

1. Add AI Agent or Basic LLM Chain.
2. Add DeepSeek AI Chat Model on the Chat Model port (AI connection, not main).
3. Select model (default deepseek-v4-flash).
4. Prefer Thinking Mode → Disabled when using tools or strict output parsers.
5. Run the workflow.

Expected sub-node UX

| During call | After call | |---|---| | Edge glow on the DeepSeek model node | Panel shows generations, tokens, cost estimate, latency, model meta |

Panel fields after a successful call (v0.1.3+):

| Field | Contents | |---|---| | response.generations | Model text (+ generationInfo / finish reason when available) | | tokenUsage | prompt / completion / total; cache hit/miss & reasoning when API provides them; source: api \| estimate | | costEstimate | USD breakdown (input cache hit/miss, output, total) using documented list rates | | meta | model, thinkingMode, durationMs, finishReasons, baseURL |

Cost is an estimate (pricing table snapshot in deepseekPricing.ts), not a billing invoice. When cache split is unknown, prompt tokens are priced at cache-miss (upper bound).

If the panel is blank or there is no glow, rebuild 0.1.3+, redeploy dist/, restart n8n.

Suggested tests

| Scenario | Thinking | Expected | |---|---|---| | Basic LLM Chain, short Q&A, no output parser | Disabled | Answer + model panel I/O | | Basic LLM Chain | Enabled | Answer; CoT at API level | | AI Agent without tools | Either | Completes; glow on model | | AI Agent with tools | Disabled | Tool loop should work | | AI Agent with tools | Enabled | May 400 if n8n strips reasoning_content |

Node parameters

Model

• Default: deepseek-v4-flash
• Also listed: deepseek-v4-pro
• Dropdown can load live models from the API when credentials work

Thinking Mode

| Value | Behavior | |---|---| | Disabled (default) | Sends thinking: { type: "disabled" } | | Enabled | Sends thinking: { type: "enabled" } + optional reasoning effort |

Reasoning Effort

Visible only when thinking is Enabled: high (default) or max.

Options

| Option | Default | Notes | |---|---|---| | Sampling Temperature | 0.7 | Omitted when thinking is enabled | | Top P | 1 | Omitted when thinking is enabled | | Maximum Number of Tokens | -1 | -1 omits the field; UI max 384000 | | Response Format | Text | json_object requires the word json in the prompt | | Timeout | 360000 ms | | | Max Retries | 2 | | | Stop Sequences | empty | Optional single stop string | | User ID | empty | Maps to DeepSeek user_id (no PII) |

Architecture (implementation)

AI Agent / Basic LLM Chain
        ▲  ai_languageModel
DeepSeek AI Chat Model  ──supplyData──► ChatOpenAI (prefer host @langchain/openai)
        │                                  + callbacks: N8nLlmTracing / fallback
        ▼
POST {baseURL}/chat/completions

| File | Role | |---|---| | nodes/LmChatDeepSeekAi/LmChatDeepSeekAi.node.ts | Node description + supplyData | | nodes/LmChatDeepSeekAi/loadChatOpenAI.ts | Resolve host ChatOpenAI | | nodes/LmChatDeepSeekAi/hostModules.ts | Host require.resolve roots (Docker/global/n8n) | | nodes/LmChatDeepSeekAi/DeepSeekLlmTracing.ts | Official tracing if present, else fallback addInputData / addOutputData | | credentials/DeepSeekAiChatApi.credentials.ts | API key + base URL + test |

Troubleshooting

Failed to parse agent steps (Basic LLM Chain)

1. Dual LangChain packages — fixed in 0.1.1+ by loading host @langchain/openai. Redeploy and restart.
2. Chain v1.9 + Structured Output Parser / “Require Specific Output Format” — known n8n bug (#29903). Workarounds: disable parser, pin chain < 1.9, upgrade n8n (~2.30+ reports fix), or parse JSON in a second step.

No output data / no running glow on the model node

Fixed in 0.1.2+ via LLM tracing callbacks (cost/latency in 0.1.3+). Rebuild, redeploy, restart. Confirm version in package.json is 0.1.3+.

Thinking on and off both fail

Usually chain/parser dual-package or n8n chain bug—not the thinking toggle. Test with thinking Disabled and no output parser.

401 / 402 / 429

DeepSeek auth, balance, or rate limits—check credential test and error codes.

Limitations

1. Self-hosted first (LangChain runtime deps).
2. Thinking + tools may 400 if n8n does not preserve reasoning_content.
3. Chat Model only—no Phase 2 action node yet.
4. No audio transcription (not in DeepSeek chat API).

Development

npm install
npm run build   # tsc → dist/ + copy deepseek.png
npm run lint    # tsc --noEmit
npm run dev     # tsc --watch

Repository layout

n8n-nodes-deepseek-ai-chat/
├── credentials/
│   └── DeepSeekAiChatApi.credentials.ts
├── nodes/
│   └── LmChatDeepSeekAi/
│       ├── LmChatDeepSeekAi.node.ts
│       ├── loadChatOpenAI.ts
│       ├── hostModules.ts
│       ├── DeepSeekLlmTracing.ts
│       ├── deepseekPricing.ts
│       └── deepseek.png
├── dist/                    # build output (gitignored; what n8n loads)
├── package.json
├── package-lock.json
├── tsconfig.json
├── .gitignore
├── README.md
├── AGENTS.md
└── PROJECT_PLAN.md

Git

• Branch: main
• node_modules/, dist/, .env* are gitignored
• Bump package.json version when releasing

Publish to npm (when ready)

npm run build
npm login
npm publish --access public

Confirm name still free: npm view n8n-nodes-deepseek-ai-chat.

References

• DeepSeek API
• Create chat completion
• Thinking mode
• Models & pricing
• n8n community nodes
• Built-in DeepSeek Chat Model

License

MIT