The AI Administrator

A full tool-using AI harness with guard-rails: debug servers in plain English and have it fix them — locally or on your remote machines over SSH — plus on-demand skills and agents you build yourself. Secrets stay local and are redacted from anything sent to a cloud model.

Not a chatbot — a full AI harness

The AI Administrator is a tool-using agent wired directly into FrontierStack. Instead of just talking, it can read logs, probe ports, check certificates, query databases, edit config, reload web servers, manage users and firewall rules, take backups and more — around 70 built-in tools in all. You drive it in plain English; it decides which tools to run and chains them together to get to an answer or a fix.

Debug in plain English, then let it fix

Ask the way you’d ask a colleague — “why is the website throwing 502 errors?” — and it investigates: tails the logs, checks the PHP-FPM pool, looks at memory and ports, and explains what it found. When it proposes a fix (restart the pool, raise pm.max_children so it won’t recur), it doesn’t just do it — it shows you exactly what it will run on an approval card, and only acts once you say yes. Read-only diagnostics run on their own; anything that changes the system is gated.

It works on your remote servers too

The same assistant operates across your fleet over SSH. Point it at a linked server and it runs the same read-only test → minimal fix → re-test loop there — updating packages, reloading Nginx/Apache, rotating a key, freeing a full disk — using key-based SSH, with no passwords ever placed in a command. So you can triage and repair a box you’re not sitting in front of, by describing the problem.

Guard-rails, by design

Approval-gated changes — every mutating action needs an explicit OK; read-only tools return immediately. Genuinely risky actions — exposing a service to the internet, opening a router/firewall port, registering a domain (real money), rebooting a host, installing software, overwriting server files, running data-changing SQL — trigger a big red confirmation that spells out exactly what is about to happen; nothing runs until you approve.
You’re told when you’re protected — when a guard-rail quietly shields you (a secret scrubbed, a change blocked in read-only mode, a send you cancelled), a green note says so — the scary-looking moments are shown to be safe.
Secrets never reach the model — passwords and keys live in a local .env vault; the AI only sees their names and writes scripts that reference $NAME, with the real values injected at run time (locally or over SSH). As a safety net, anything bound for a cloud engine is double-scrubbed: every credential the app has stored (its Keychain and your linked password managers) is removed by exact match, plus secret-shaped patterns (tokens, keys, private keys, connection strings) and usernames.
Optional local-AI prompt screening — if you have a local model (Apple on-device or Ollama), a second pair of eyes runs on your Mac to catch secrets, PII or internal hostnames the patterns might miss before anything reaches a cloud API — and can warn & confirm, auto-redact, or block. The check itself never leaves the machine, and sits on top of the always-on redaction, never instead of it. The option appears only when a local model is available.
Your choice of engine — run it on any major cloud AI platform, a local Ollama or LM Studio model, or Apple’s on-device Foundation Models framework. Local engines keep everything on your Mac, with no redaction needed.
It remembers your preferences — durable rules like “my messaging app is LINE” or “don’t make changes on weekends” persist across sessions and shape how it acts.

On-demand skills

For bigger jobs it loads multi-step skills — a fleet-wide security audit, a malware analysis pass, SSH hardening — that walk a checklist of tools and hand you a report.

Build your own agents

Beyond the interactive assistant, you can create agents inside FrontierStack: goal-driven loops tuned to your use-case. Give one a goal (“watch this endpoint and restart it if it stalls”, “every morning, reconcile inventory and flag anything off”), choose whether it may make changes, set a step budget, and run it once or on a schedule. It uses the same guarded tools to monitor things or do real work for you, streaming every step to a transcript you can stop at any time.

Drive it from Claude Code & other harnesses (MCP)

FrontierStack also runs as its own MCP server. Switch it on and an external AI harness — Claude Code, Claude Desktop, Cursor, Codex or any MCP client — can call FrontierStack’s tools directly. Those calls route to the same guarded executors as the built-in assistant and act on live app state, so your coding agent can inspect and operate your servers through FrontierStack. The endpoint binds to localhost behind a required bearer token (optionally exposed to your LAN), every request is logged, and it stays read-only until you opt in to changes and scripts — with the same in-app approval card in front of anything that mutates state.

Because the harness is your own subscription tool, the AI usage rides on the plan you already pay for — Claude Code on a Claude Pro/Max plan, Cursor on its subscription, and so on. You get FrontierStack’s tools driven by the assistant you use every day, at a flat rate, with no metered per-token API bill.

Supported AI platforms

Run the built-in AI Administrator on whatever you prefer — add a key in the AI Models pane, or keep it local:

Cloud models — OpenAI, Anthropic (Claude), Google (Gemini), xAI (Grok), DeepSeek, Mistral, Groq, Together AI, OpenRouter, NVIDIA NIM, Novita, Moonshot (Kimi), Z.ai (GLM), MiniMax, Xiaomi MiMo and Merlin.
Local models — Ollama, LM Studio, Docker Model Runner, or any OpenAI-compatible local server (Apfel, LocalAI, Jan, vLLM, llama.cpp) installed and launched from the Local AI section — fully offline on your Mac with nothing redacted.
Apple Foundation Models (on-device) — Apple’s private Foundation Models framework built into macOS, with no API key and no network. Available on Apple silicon with macOS 26 and Apple Intelligence enabled — or via Apfel, which exposes the on-device model through a local OpenAI-compatible API. FrontierStack’s Apple Intelligence section configures and runs it for you.

And to drive FrontierStack from outside the app, connect any MCP harness — Claude Code, Claude Desktop, Cursor, Codex or another MCP client (above).

One monitor for every AI bill

Spread your work across providers, subscriptions and harnesses and the spend scatters. The AI Model Costs dashboard is a central monitor for all of it: flat subscriptions (ChatGPT, Claude Pro/Max, Cursor, Perplexity, T3 Chat…) and metered API token spend per provider and model, with voice/speech minutes too. It shows month-to-date totals and a per-provider breakdown against a monthly budget bar, and a billing-trouble watcher raises an Alert on out-of-credits, hard-limit, past-due or a suspended account — so a key never dies on you mid-job.

The AI Administrator, the MCP server and all remote actions are off until you enable them, and change permissions are yours to grant.

This is the web edition of FrontierStack’s in-app Help. Open the same topic inside the app for live, clickable controls.

Run it all from one Mac app.

FrontierStack installs, monitors and secures the whole stack — locally and across your fleet — from a single native macOS app.

Download FrontierStack