Which model runs is a configuration decision, not a runtime one.
NopsAI accepts nine provider values. What matters more than the list is where the choice lives: in a named profile under Git review, scoped to environments, rather than in a string a step assembles while it runs.
LLM providers inside a governed run.
How LLM providers participates once it is connected — and what the platform does with what it returns.
Nine providers, one configuration shape
Gemini, OpenAI, Anthropic, Azure OpenAI, Groq, Mistral, OpenRouter, Ollama and LM Studio are all configured as profiles with a provider, a model, an optional base URL and a credential reference. Switching provider does not change a pipeline.
Hosted and self-hosted side by side
Ollama and LM Studio take a base URL and typically no credential; the hosted providers take a credential reference. Both kinds can exist in the same deployment, letting a sensitive step stay local while another uses a stronger hosted model.
Selection at three levels
llm_profile can be set on the pipeline, on a step and on a task, with the narrowest winning. A default profile covers everything that does not state a preference.
Tuning belongs to the profile
Reasoning effort, thinking, timeout, max tokens, temperature, prompt caching and provider-specific extras are profile settings, so they are reviewed once rather than repeated across pipelines.
The controls that make it safe to leave connected.
An integration is easy to add and hard to bound. These are the parts that decide whether it stays reviewable six months later.
Environment scoping is enforced
allowed_scopes decides where a profile may be selected. An experimental model available in dev cannot be picked up by a production run, and the mismatch is a configuration error rather than a silent fallback.
Keys never live in Git
Credentials are referenced as credential://system/llm/... and resolved from the encrypted registry with versioning, rotation and consumer access logs. Configuration repositories stay safe to review widely.
Content sharing is decided upstream
llm_content_sharing, llm_content_include and llm_content_ignore decide what may be sent before any request is built. Approving a provider is not the same as approving it to see the whole workspace.
Usage is attributed per subject
AI usage events record the profile and the effective subject, so spend and model use can be attributed to a person and a workflow rather than to one shared account.
Owned by Git, reviewed like code.
Configuration is a reviewable file rather than a form someone filled in once. Credentials appear as references; the values live in the encrypted registry.
default_profile: standard
profiles:
- name: fast
provider: gemini
model: gemini-2.5-flash
credential_ref: credential://system/llm/gemini-fast
allowed_scopes: ["dev", "prod"]
- name: reasoning
provider: lmstudio
model: google/gemma-4-26b-a4b
base_url: http://lmstudio:1234
reasoning: high
allowed_scopes: ["dev"]
- name: claude-review
provider: anthropic
model: claude-sonnet-4-6
credential_ref: credential://system/llm/anthropic-review
max_tokens: 4096
allowed_scopes: ["prod"]Workflows where LLM providers shows up.
The same integration, seen from the workflow rather than from the connection.
Sprint planning for the next two weeks
Planning starts with manual research instead of decisions. Blockers surface late and half-ready stories enter the sprint.
Production incident at 2 a.m.
At 2 a.m. the hard part is finding what changed and deciding whether rollback is safe — fast, without losing the evidence trail.
Preparing a release
Scope mismatch, unmerged PRs, failed gates and an outdated runbook can all cause a late surprise.
What teams ask about LLM providers.
Which providers does NopsAI support?
Gemini, OpenAI, Anthropic, Azure OpenAI, Groq, Mistral, OpenRouter, Ollama and LM Studio. Each is a first-class provider value in configuration rather than a generic OpenAI-compatible shim.
Can one pipeline use more than one provider?
Yes. Because llm_profile applies at pipeline, step and task level, a single run can route a sensitive step to a local model and a summarization step to a hosted one.
What stops a team using an unapproved model?
Pipelines reference profiles by name and cannot define a provider or endpoint inline. Adding a profile is a reviewed change to system configuration, and allowed_scopes then limits where it can be used.
Map your LLM providers workflow end to end.
Bring one real workflow. We will map its trigger, tools, approvals, runtime boundary and evidence requirements with you.