Help Other AI Server

Help: Other AI Server

Other AI server lets LrGeniusAI use any server that speaks the OpenAI chat API. You only need it if the AI should run somewhere else than on this computer — for example:

• OpenRouter — one key for hundreds of cloud models, including free ones.
• llama.cpp's llama-server, vLLM or LiteLLM on your own machine or network.
• LM Studio or Ollama running on another computer (a desktop with a big GPU, a NAS, …).

You do not need it for:

• the built-in local model (AI model (on this computer) in Plug-in Manager — see Local AI Models);
• Ollama or LM Studio running on this computer — LrGeniusAI finds them automatically at their default address;
• OpenAI, Google Gemini and Anthropic (Claude) — they have their own key fields.

Anthropic does not work here. Anthropic offers an OpenAI-compatible address (https://api.anthropic.com/v1), but it only answers chat requests: the model list behind it is Anthropic's own and fails with "HTTP 400 (anthropic-version: header is required)", and it ignores the answer format LrGeniusAI asks for. Enter your key in the Anthropic key field instead — or, if you want Claude through OpenRouter, use OpenRouter here.

Setting it up

File → Plug-in Manager → LrGeniusAI → Optional AI providers:

1. Other AI server — the server's address.
2. API key — only if the server needs one. Leave it empty otherwise.

The line under the two fields checks the server as soon as you leave a field: it shows how many of the server's models can read photos (OpenRouter: 212 models that can read photos), or what is wrong. The models then appear in every task's AI Model list under the server's name — OpenRouter · google/gemini-2.5-flash, or the host and port for your own server (192.168.1.20:8080 · gemma-3-12b-it).

Only models that can take a photo are listed: text-only and embedding models are left out where the server says which is which.

Addresses

Paste the address the server's documentation gives. LrGeniusAI is lenient about the form:

You enter	Used as
https://openrouter.ai/api/v1	as is
openrouter.ai/api/v1	https://openrouter.ai/api/v1 — no scheme means https for internet addresses…
192.168.1.20:8080	http://192.168.1.20:8080/v1 — …and http on your network or with a port; no path means /v1
http://gpu-box.local:1234/v1/chat/completions	http://gpu-box.local:1234/v1 — a pasted endpoint is cut back

Examples:

Server	Address	API key
OpenRouter	https://openrouter.ai/api/v1	your OpenRouter key
llama.cpp llama-server	http://<host>:8080	only if started with --api-key
LiteLLM proxy	http://<host>:4000	the proxy key, if a master key is set
vLLM	http://<host>:8000	only if started with --api-key
LM Studio on another computer	http://<host>:1234	only if its server requires a token
Ollama on another computer	http://<host>:11434	none

Things to know

• The server gets your photos. Every photo you analyse is sent to it, together with the context you chose to send (keywords, GPS, folder names). For a cloud service like OpenRouter that means the photo leaves your computer, and the provider's terms apply.
• Costs. Paid models on OpenRouter cost credits per photo, like OpenAI or Gemini. OpenRouter's :free models cost nothing but are rate limited.
• One photo at a time. Requests to your own server are sent one after the other, as for Ollama and LM Studio: most such servers run a single model that would process them one by one anyway.
• llama-server needs its vision projector. Start it with --mmproj (or -hf, which downloads it), or every photo fails with "cannot read photos".
• Ollama on another computer is reached through its OpenAI-compatible API, which cannot set the context size per request. If long prompts get cut off, raise it on that computer, e.g. OLLAMA_CONTEXT_LENGTH=8192 ollama serve.
• Answer format. LrGeniusAI asks the server to hold the model to a JSON schema. A server that cannot do that is asked again with a looser format (it remembers this per model, so only the first photo pays for the extra request), and the photos answered that way carry a warning saying so. The answers are still checked, but a weaker model may leave a field out more often — that shows up as a warning on the photo, as with any other provider.

When it does not work

The status line and the task's error dialog say what went wrong, in words:

Message	What to do
Could not reach …	Check the address and that the server is running and reachable from this computer (firewall, VPN).
… rejected the API key / … needs an API key	Copy the key again from the provider's page and paste it in — the field shows only dots, so you cannot check it by eye.
… has no OpenAI-compatible API at this address	The address usually ends in /v1 — check the server's documentation.
… does not know the model …	The model was removed or renamed; pick another in the AI Model list.
… cannot read photos	Pick a vision model; for llama-server, add --mmproj.
… is rate limiting requests	Wait a little, or use a model without a free-tier limit.
… no credit left on the account	Top up the account with the provider.

Moved from the old Ollama / LM Studio fields

Earlier versions had separate address fields for Ollama and LM Studio. If you had changed one of them (the app ran on another computer), that address was moved here once, and your model choice with it; the Plug-in Manager shows a note the first time it opens. If you had changed both, only the one you were using was moved — enter the other here if you need it instead.