// docs

S.E.N.T.R.I

Server Engine Neural Task Runner & Intelligence — a fully agentic AI that plans, calls skills, runs runbooks, and acts on your infrastructure.

How it works

Give S.E.N.T.R.I a goal in plain language. It breaks the goal into steps, calls the appropriate skills, runs your runbooks, checks results, and reports back — asking for approval where you require it. You bring the model; S.E.N.T.R.I brings the tools and the guardrails.

Bring your own model

Choose the provider that fits your policy and budget:

OpenRouter

Bring your own OpenRouter API key (BYOK) and pick any model OpenRouter offers. ServerEngine shows your remaining budget in the chat so you always know your spend.

  1. 1

    Get a key

    Create an API key at openrouter.ai and paste it into the S.E.N.T.R.I settings.
  2. 2

    Pick a model

    Enter any OpenRouter model ID; set the reasoning effort if the model supports it.
  3. 3

    Star the ones you use

    Mark your go-to models as favorites, then filter the list to show only those — no scrolling past hundreds of models to find the two you actually run.

Claude (Claude Code Integration)

Run S.E.N.T.R.I on Claude through the Claude Code Integration— ServerEngine drives the Claude Code CLI in the background as the model. There's no separate API key: it reuses your Claude Code login. Because the CLI is the backend, whatever it's configured for works out of the box — a Claude subscription (Pro/Max), an API key, or enterprise routes like Amazon Bedrock and LiteLLM gateway configs. No extra setup in ServerEngine.

Prerequisite

Install and sign in to the Claude Code CLI first (available from claude.com/claude-code). ServerEngine detects it automatically. Choose a model alias such as opus-4.8, sonnet-5, or haiku-4.5.

Ollama (fully local)

Point S.E.N.T.R.I at a local Ollama endpoint (for example http://localhost:11434) and pick a model to keep everything on-premises — no data leaves your network.

What reaches your model provider

S.E.N.T.R.I has to send the model what it is reasoning about — your prompt, the host names in the current scope, and the output of the skills it calls. With OpenRouter or Claude Code that context goes to the provider you chose. Stored credentials are never part of it: they are substituted into scripts by the service at execution time, not by the agent. Choose Ollama if nothing may leave the network at all.

Interactive and unattended

S.E.N.T.R.I runs two ways, and the difference matters for your guardrails:

  • Interactive — you chat with the agent in the console. Skills marked approval-required pause and wait for you to Accept or Decline.
  • Unattended — a scheduled prompt or an API-triggered Prompt job, executed headlessly by the automation service. No one is there to approve, so approval-required skills are skipped; the scope and your enabled skills are the boundary.

Both write the full conversation and every action to the Logs under a job ID.

Chat features

  • Image analysis — attach PNG, JPG, or WebP images (OpenRouter and Ollama providers) and ask S.E.N.T.R.I about them.
  • Code blocks — responses render code with syntax highlighting; internal skill-call JSON is hidden from the transcript.
  • Context stats — a live token/context readout and, for OpenRouter, your remaining budget.

Next: the skills catalog, approvals, and scopes →