Use your chat subscriptions
Supported subscription paths can include ChatGPT, Gemini, xAI Grok, and Kimi. A subscription connection can avoid a separate API key when the provider and current Jarvis plan support that method.
Your AI
Use the AI you choose.
Connect chat subscriptions, API keys, or your own models. Change the AI without changing how your team works in ERPNext.
Your options
Three ways to connect.
Use provider API keys
Native and compatible connections can cover providers such as OpenAI, Anthropic, Google Gemini, Mistral, Groq, Together AI, DeepSeek, Moonshot Kimi, xAI Grok, GLM and Z.ai, and OpenRouter.
Use your own models
OpenAI-compatible endpoints, Ollama, vLLM, Qwen through a compatible endpoint, and other approved private services can fit deployments that need local model control.
Provider names describe supported connection families, not a permanent guarantee. Providers can change authentication, external access, model names, or terms. The live AI model settings show what the workspace can connect now.
Why does choice matter?
AI services differ in cost, context limits, speed, availability, data terms, and the quality of particular business tasks. Locking the entire ERPNext workflow to one provider can make a temporary model choice feel like a permanent system decision.
Jarvis keeps the business workflow around the model: the current user, permitted ERPNext records, source links, prepared work, review, and reusable routines. An administrator can change an approved connection without teaching the team a different operational process.
How do you control AI cost?
Yes, but cost control begins with a bounded task and visible usage, not only with choosing the cheapest model. A company can start with an existing supported subscription, use API pricing for workloads that fit it, or run approved local models where the operating cost and maintenance make sense.
Begin with one task whose value is obvious, such as an overdue invoice review, a supplier bill queue, or a weekly owner dashboard. Compare the time removed, the quality of the checked result, and the AI usage before expanding.
What if a model fails?
An administrator can order approved models so Jarvis has another connection when the preferred one is unavailable. A longer list is not automatically safer. Every provider adds its own credentials, cost, data terms, rate limits, and behavior.
Fallback protects availability, not identical judgment. Review important work after a model changes because wording, tool choices, and interpretation can differ.
Who manages the API keys?
Only authorized administrators should add, reconnect, order, disconnect, or remove model connections. Keys and account credentials belong only in the connection settings. They should never be pasted into Chat, Wiki, Business Notes, or Support.
Before enabling a provider, the company should review the provider's current data handling, regional availability, retention, commercial terms, and whether the intended external connection is allowed by that provider.
Choose well
Choose for the task, not the logo.
Name the first task
Decide whether the first value is chat, document preparation, dashboards, scheduled checks, or a repeated routine.
Approve the provider
Review the provider, connection method, data terms, cost owner, and administrator responsible for credentials.
Ask before making changes
Ask one ordinary question with a normal user role and inspect both the answer and source records.
Check value and cost
Compare time saved and usable results against the actual usage shown for the workspace.
Add a backup when needed
Add another model only when its availability, capability, or economics solve a real risk.
Check again after a change
Treat a new model or connection order as a workflow change that deserves a checked run.
Change the AI, not the work.
Connect the approved service, ask one real ERPNext question, and expand only after the checked result earns it.