Add support for llama.cpp

[vidual-labs]

It uses open ai protocols but I would need to be able to enter the endpoint url myself

[ghdoca]

I fully support that request! Please adapt it for use exclusively with local models: llama-server and whisper.cpp. No cloud APIs needed.

[ghdoca]

After a day of hassle I managed to run it locally with llama-server (qwen-coder-next).
Replaced Claude with OpenAI nodes, remember NOT to use ResponsesAPI.
Only tested text input through telegram and MCP weather, works. Heartbeat does not work, as the Claude Haiku call is embedded in js.
Also webhook was a major problem as I'm not using any standard port but 88.

All in all ok, updates of course will be a problem...

[freddy-schuetz]

Thank for your request! n8n-claw already works with llama.cpp, Ollama, and any OpenAI-compatible LLM endpoint! Thanks to @ghdoca for confirming this works in practice.

How to switch

Replace the Anthropic LLM node with an OpenAI Chat Model node in these 3 workflows:

Workflow	Node to replace	Default
🤖 n8n-claw Agent	"Claude" (LLM sub-node)	Claude Sonnet
🏗️ MCP Builder	"Anthropic Chat Model" (2 nodes)	Claude Opus
🧠 Sub-Agent Runner	"Claude" (LLM sub-node)	Claude Sonnet

Steps per workflow:

1. Open in n8n → delete the Anthropic LLM node
2. Add an OpenAI Chat Model node → connect it to the AI Agent node
3. Set Base URL to your endpoint (e.g. http://localhost:8080/v1 for llama.cpp)
4. Disable Responses API in the OpenAI node (important!)
5. Create a credential with your endpoint + any API key string

Known limitations

• Heartbeat + Memory Consolidation still use hardcoded Anthropic API calls in JavaScript code nodes — these won't work with alternative providers yet. Deactivate them if running without an Anthropic key. Making these provider-agnostic is on our roadmap.
• setup.sh requires an Anthropic API key (the prompt doesn't accept empty input). Enter any placeholder value like sk-dummy if you're only using local models — the key is only used by Heartbeat + Memory Consolidation.
• We just merged a fix (#13) that makes the expert agent system more robust with non-Anthropic LLMs (some models prefix tool call JSON with emoji text — this is now handled gracefully).

I've added a full Alternative LLM Providers section to the README with all details.

If you test with other providers, please share your experience here — I'll add it to the compatibility table!