I’m excited to share that I’ve just released a new n8n community node for DeepSeek that finally makes the DeepSeek Chat Model play perfectly with the AI Agent node and Tool Calling.
What this node does
-
Provides a dedicated DeepSeek Chat Model (Corrected) community node for n8n self-hosted
-
Runs smoothly with AI Agent Tool Calling
-
Completely avoids the annoying error message like:
“The reasoning_content in the thinking mode must be passed back to the API.”
Instead of fighting with thinking mode and reasoning_content handling on each turn, you can simply disable Thinking Mode on this node and let the AI Agent handle tool calls as expected.
Node on npm: n8n-nodes-deepseek-chat-model
Current limitations
Right now, this community node can only be installed and used on n8n self-host (it is not available on n8n Cloud because it’s an unverified community node).
How to install and use
Follow these steps on your self-hosted n8n instance:
-
Go to Settings → Community Nodes.
-
Click Install, paste the package name: n8n-nodes-deepseek-chat-model, then click Install.
-
Restart your n8n container so the whole system (especially in queue mode) properly loads the node.
-
In your workflow, open the AI Agent node and select DeepSeek Chat Model (Corrected) as the Chat Model.
-
Turn Thinking Mode off on this node.
-
Connect your tools as usual and start using the AI Agent with Tool Calling.
With this setup, DeepSeek Chat Model (Corrected) will run reliably with tool calls, without throwing reasoning_content errors, and you get a clean, stable agent experience on n8n.
2個讚
@Anshul_Namdev what do you think about this??
Hi @nguyenthieutoan, thank you so much for putting this together!
Just wanted to share my use case in case it helps others out there. I run a self-hosted n8n instance (via Docker) using a WhatsApp AI Agent that relies heavily on Tool Calling to execute an SQL query tool to check for availability in a MySQL database.
The Problem: I was constantly hitting the 400 reasoning_content error when DeepSeek tried to execute the tool. I temporarily switched to Gemini to keep the bot running in production, but Gemini’s tool-calling consistency was nowhere near DeepSeek’s precision for my specific use case.
The Solution: I installed your n8n-nodes-deepseek-chat-model node, selected the DeepSeek Chat Model (Preserved Reasoning), created a new API credential, and strictly disabled the Thinking Mode. It worked perfectly! The agent is now calling the SQL availability tool flawlessly again without breaking the workflow.
Quick tips for other self-hosted users deploying this:
-
Restart the container: After installing the node via the UI, don’t forget to restart your n8n docker container (e.g., docker compose restart), otherwise, the node might not show up in your workflow editor due to cache.
-
Recreate Credentials: Since this is a community node, your existing official DeepSeek credentials won’t port over automatically. You just need to click ‘Create new credential’ on the new node and paste your API key again.
Thanks again for saving our production flows while we wait for the official fix (AI-2422)!
2個讚
@W3_Servicos So glad it helped your production setup!
I’m also building and running SaaS in production, so I really feel your pain when an AI/LLM change suddenly breaks a live workflow.
If you’re open to it, I’d love to chat more about your production products and your WhatsApp AI Agent setup, maybe we can swap feedback on each other’s products or even find a way to collaborate.
2個讚
Thank you!!! help a lot!!!
We can definitely chat about it! I’d be happy to share whatever might be helpful, though my perspective is somewhat limited to my local business sector. I probably won’t have much to offer in terms of workflow optimization or coding, given your extensive expertise in those areas. However, from a networking perspective, I do have connections with high-level SaaS resellers in my region, so some good business opportunities could certainly arise from our interaction
att. Pedro
atendimento@w3vistorias.com.br
Hi @nguyenthieutoan,
Just a quick follow-up to share a technical feedback on how we evolved our production setup after using your community node.
While your node successfully resolved the 400 reasoning_content error, we later noticed a parsing mismatch in LangChain that caused raw DSML/XML tags (like <｜｜DSML｜｜tool_calls>) to leak into the production WhatsApp chat.
To solve this leakage while keeping DeepSeek, we migrated to the official OpenAI Chat Model node acting as a bridge:
Since DeepSeek is OpenAI-compatible, LangChain now handles tool-calling natively and invisibly without leaking code. Thanks again for providing the initial fix when we needed it!