Docs
Philosophy

Infrastructure on Autopilot

AI that monitors, manages, and acts on your behalf. Not a chatbot — an agent that uses your tools, visibly.

Transparent · No black box Ghost Mode · Watch AI work Automation · Rules you control

What We Believe

Most AI assistants are chatbots pretending to be helpful. You ask them to do something, they say "Sure!", then disappear into a black box. Maybe it worked. Maybe it didn't. You won't know until something breaks.

We think that's backwards.

AI should be more transparent than clicking buttons yourself — not less. When you tell Caveman to deploy your service, you should see exactly what it's doing: which form it's filling, what values it chose, where it paused because it needs your input.

No Magic

Everything AI does is visible. No hidden API calls, no secret actions. You see the same forms AI fills.

Same Code Path

AI uses the exact same UI you do. Same validation, same permissions, same everything. No backdoors.

User Learns

Watch AI do it once, do it yourself next time. AI teaches you the system by using it in front of you.

Transparency Over Magic

Traditional AI
"I'll deploy that for you"
...silence...
"Done!"
What actually happened?
Caveman AI
Opens /services/new
Fills: Name = "gyr", Host = Fly.io
"Which region?" [iad] [sjc] [ams]
Continues with your choice

You're not trusting a black box. You're watching an expert use your tools — and learning how to do it yourself.

Three Ways AI Helps

Reactive You ask

AI shows plan, fills forms, you confirm key steps

deploy gyr to fly
Advisory AI notices

AI suggests action, you decide what to do

Memory 40% higher than normal
Autonomous Rule matches

AI acts automatically, you get notified

CPU > 90% for 10 min

You control how much autonomy AI has. Reactive requires your confirmation for everything. Advisory shows alerts but waits. Autonomous follows rules you pre-approve.

The Action Bar

One input does everything: search, commands, natural language. No mode switching. History shows actions, not conversation.

deploy gyr to fly 32gb

CPU at 98% on api-gateway for 3h

Scale 3→5? Or write better code

Activity

Auto-scaled api-gateway 3→5

Rule: "High Load" triggered • 5m ago

Deployed gyr v2.1.3 → fly/iad

15m ago

Warning: Memory creeping up on ml-inference

Restarted automatically • 1h ago

Ghost Mode

When AI performs a complex action, it shows each step in real-time. Like watching an expert fill out forms — except you can intervene at any point.

DEPLOYING GYR → FLY.IO
Step 1: Create Service
Step 2: Environment Variables
DATABASE_URL From Secrets ✓
REDIS_URL From Secrets ✓
API_KEY ? Enter value
API_KEY not in your secrets. Enter value or add to secrets?
Step 3: Specs & Scaling
Step 4: Review & Deploy

Automation Rules

Rules are pre-approved workflows. When conditions match, AI acts automatically — no confirmation needed, because you already approved it.

Create rules with natural language:
if cpu > 90% for 10min then scale up 2 machines
Got it. Rule "High CPU → Scale Up" is now active.

Scale Up — High Load

If job queue > 1,000 → Scale to 5 machines

12× this week

Scale Down — Low Traffic

If job queue < 3 → Scale to 1 machine

8× this week

Hot Potato

If CPU > 90% for 10min → Add 2 cores

3× this week

AI Monitoring

AI continuously watches your services and sees what dashboards can't: patterns, anomalies, correlations, predictions.

Anomaly

Memory 40% higher than normal

Possible leak in api-prod since deploy #847

Root Cause

Latency spike correlates with deploy

Response time increased 3× after deploy #847

Proactive

You'll hit memory limits in 3 days

Based on current growth rate

Cost

staging-worker is idle

4 requests in 7 days, costing $23/month

Key Principles

1

One input does everything — no mode switching

2

No chat bubbles — show action cards, not messages

3

AI is invisible until it has something to say

4

History = actions performed, not conversation

5

Ghost Mode — AI shows its work, user can intervene

6

Rules = pre-approved workflows with constraints

7

Transparency > Magic — users learn by watching

The bottom line: AI should make you better at managing your infrastructure, not dependent on it. When you watch Caveman work, you're learning. When you set up rules, you're teaching. It's a partnership, not a black box.