Alma by Olivares AI

Olivares AI is a self-hosted, vendor-neutral platform for integrating, managing and securing enterprise AI on the infrastructure you already run. It discovers agents, sessions, models and MCP servers, then builds an access map of what each one can read or write. The goal is to give platform, security and operations teams ground truth they can use for policy, cost and evidence.

A gateway focuses on model calls: routing, keys, budgets, caching and sometimes content filters. LLM observability focuses on traces, spans and prompt-level debugging. Olivares sits around the estate instead, mapping agents to the resources they actually touch and comparing permitted access with observed behavior. Those tools can stay in place; Olivares adds the infrastructure-level governance and evidence layer beside them.

Inference against a closed frontier model — Claude, GPT, Gemini — is not self-hostable, so that request still reaches the vendor or a supported cloud provider account you configure. What is self-hosted is Olivares itself: discovery, access map, audit ledger, policy state and governance data run in your infrastructure. Air-gapped means that governance and observation layer can operate without calling home. For models you genuinely self-host, inference can stay local too.

No. The access map, identity, policy, FinOps and evidence layers are vendor-neutral by construction — designed around what agents touch, not around one model vendor — and the model catalog spans Anthropic, GLM, OpenAI and Google alongside anything you self-host on vLLM or Ollama. Integration depth is uneven and we would rather say so than blur it: Claude and Claude Code are the deepest surface today, with OpenAI Codex and Grok Build governed through the same console. Deepest is not the same as only.

No. Olivares complements Claude; it does not replace it, resell it or broker a subscription credential. Your model relationship remains between you and the model provider. Olivares is the governance and observation layer around the agents and sessions that use those models.

Olivares governance data stays in your perimeter: the access map, audit ledger, inventory and policy state live in your storage. The platform records relationships and outcomes rather than prompt payloads, result rows, secrets or personal data. There is no mandatory telemetry home. The honest exception is model inference itself: hosted models such as Claude still receive the requests you choose to send to them.

The discovery and access-map path is out of band. It reads audit logs, OpenTelemetry and source-native signals rather than sitting in every agent data path, so a collector failure degrades visibility rather than production traffic. Enforcement only applies where you deliberately connect a governed gate, such as a Claude Code hook or another provisioned actuation surface. Those gates are designed to fail closed rather than show a false green.

Trust each edge exactly as far as its evidence says. The map distinguishes read from read/write only where the source can prove that distinction, and it carries confidence about attribution and coverage. If a shared service account hides the agent identity or a source cannot prove read versus write, the product marks that uncertainty instead of guessing. Clean evidence is actionable; unknown or approximate evidence is a prompt to improve instrumentation.