What happened, what we decided, why, and how to build yours
AZ Rollin — March 3, 2026
Here's what went down in one session:
1 Installed Docker Desktop — the "building" where our AI team lives.
2 Installed Ollama — a free AI brain that runs directly on my Mac (M3 Max, 48GB RAM). No internet needed. No monthly fees. Zero cost.
3 Deployed OpenClaw — the "office" where AI agents work. Runs 24/7 in Docker.
4 Created MAYA & OLIVER — my first two agents. MAYA handles priority tasks (named after my niece, age 8). OLIVER handles support tasks (named after my nephew, age 5). They cross-check each other on important decisions.
5 Secured the dashboard — changed to a random port, locked it to localhost-only, added token authentication + device pairing.
6 Hit a wall — the free AI brain (Llama 3.1 8B) was too small for agent work. It broke in weird ways...
Llama 3.1 8B (the small free model) couldn't handle agent work.
Why? Running an AI agent is WAY harder than just chatting. The agent needs to:
An 8-billion parameter model just doesn't have enough "brain power" for that. It's like asking an intern to be CEO.
The lesson: Free local AI is amazing for simple tasks (drafting, summarizing, quick research). But for running a team of agents? You need more power.
We researched every option. Here's what we found:
$90-300/month
Smartest option
Also most expensive
~$9/month
95% cheaper than GPT-4o
Nearly as smart
Free forever
Needs 40GB+ RAM
Slower, offline backup
Why DeepSeek?
Bottom line: When you're bootstrapping, spending $9/month instead of $150/month means you survive longer.
The backup plan: Ollama with Llama 3.1 70B (40GB model) is still downloading. Once it finishes, it becomes our free offline backup — agents can still work when the internet goes down or if we want to save every penny.
Two are live. Nine more are designed and ready to deploy when needed.
ACTIVE Priority tasks, content pipeline, revenue tracking, daily action-log review. Makes the big calls. Always cross-checks with OLIVER before approving anything major.
ACTIVE Light tasks, cross-checks MAYA's decisions, quality control, security watchdog. The safety net. If MAYA says "go," OLIVER double-checks before we actually go.
PLANNED Web searches, competitor analysis, trend tracking, market research. Reports findings to MAYA.
PLANNED Episode scripts, YouTube descriptions, social captions, email drafts. Never posts — always drafts for AZ's approval.
PLANNED Revenue tracking, expense logging, affiliate commission monitoring, financial forecasting. Reports to MAYA.
PLANNED Social media scheduling recommendations, hashtag research, engagement analytics, cross-platform strategy. Draft-only.
PLANNED Builds custom tools, scripts, automations, web pages. Gets the smartest brain (Claude Sonnet) because coding needs precision.
PLANNED Analyzes video transcripts, suggests clip timestamps, writes shorts descriptions. Works with ECHO on content pipeline.
PLANNED Launch monitor, YouTube analytics, social metrics, trend detection. Alerts MAYA when something needs attention.
PLANNED Translates content into Azerbaijani, Russian, Turkish. Adapts tone for each culture, not just word-for-word translation.
PLANNED Monitors comments/feedback, gauges audience sentiment, suggests content adjustments. The "finger on the pulse."
Not every agent needs the same power. That's how you keep costs low. It's like hiring: you don't pay intern rates for a CEO, and you don't pay CEO rates for filing paperwork.
| Agent(s) | AI Model | Why This Model | Cost |
|---|---|---|---|
| MAYA, OLIVER | DeepSeek V3 | Smart enough for management decisions. 95% cheaper than GPT-4o. | ~$0.30/day |
| ECHO, BOOST, CLIPPER, PULSE | DeepSeek V3 | Writing and content work. DeepSeek handles it well. | ~$0.15/day |
| SCOUT, RADAR, DOLLAR | DeepSeek V3 | Research and analysis. Good with data processing. | ~$0.10/day |
| BUILDER | Claude Sonnet 4.6 | Coding requires the smartest model. Sonnet is the best coder. | ~$1-3/day (when used) |
| BRIDGE | DeepSeek V3 | Translation quality is excellent in DeepSeek. | ~$0.05/day |
| ALL (backup) | Ollama / Llama 3.1 70B | Free offline fallback. When internet is down or saving money. | $0 |
Total estimated daily cost (full team running): ~$1-2/day ($30-60/month)
Compare to running everything on GPT-4o: $10-20/day ($300-600/month).
That's 85-90% savings.
1 Go to platform.deepseek.com
2 Click "Sign Up" — use email or Google account
3 Once logged in, go to API Keys (left sidebar)
4 Click "Create API Key"
5 Give it a name like openclaw-team
6 COPY THE KEY IMMEDIATELY — it starts with sk- and you can only see it once!
7 You get 5 million FREE tokens on signup — no credit card needed
Cost: $0 to start. When free tokens run out, add $5-10 credits. At ~$0.30/day, $10 lasts about a month.
1 Go to console.anthropic.com
2 Sign up or log in (this is SEPARATE from your Claude Max subscription)
3 Go to API Keys → Create one named openclaw-builder
4 Add $5 credits to start
The Llama 3.1 70B model (~40GB) is currently downloading on your Mac. Once it finishes, it becomes a free offline backup. No API key needed — it runs on your machine.
Check status: ollama list in Terminal
Once you have your DeepSeek API key, give it to Claude Code in Terminal and say:
"Here's my DeepSeek API key: sk-xxxxx. Add it to OpenClaw and switch MAYA and OLIVER to DeepSeek V3."
What Claude Code will do:
~/.openclaw/openclaw.jsonFor each new agent (SCOUT, ECHO, etc.), just tell Claude Code:
"Add SCOUT as a research agent using DeepSeek V3."
Claude Code will create the agent's workspace, personality file, and register it in the config. Takes about 2 minutes per agent.
| What | Cost | Notes |
|---|---|---|
| OpenClaw (Docker) | Free | Open source, runs on your Mac |
| Ollama + Llama models | Free | Local AI, no internet needed |
| DeepSeek V3 API | ~$9/month | 5M free tokens on signup, then ~$0.30/day |
| Anthropic API (BUILDER only) | ~$30-90/month | Only when coding. Skip until needed. |
| Claude Code (Terminal) | $0 extra | Included in Claude Max subscription |
| Telegram Bot | Free | Already running |
| Minimum (MAYA + OLIVER only) | ~$9/month |
| Full team (all 11 agents, no BUILDER) | ~$30/month |
| Full team + BUILDER (Claude Sonnet) | ~$60-120/month |
Compare: Running this on GPT-4o alone would cost $300-600/month.
| What | Details |
|---|---|
| Open Dashboard | http://localhost:YOUR_PORT |
| Start OpenClaw | cd ~/openclaw && docker compose up -d openclaw-gateway |
| Stop OpenClaw | cd ~/openclaw && docker compose down |
| Check health | cd ~/openclaw && docker compose logs --tail 10 openclaw-gateway |
| Restart after config change | cd ~/openclaw && docker compose restart openclaw-gateway |
| Talk to MAYA | Dashboard → Chat → select MAYA → New session |
| Talk to OLIVER | Dashboard → Chat → select OLIVER → New session |
| Check Ollama models | ollama list |
| Config file | ~/.openclaw/openclaw.json |
| Docker env | ~/openclaw/.env |
| MAYA's personality | ~/.openclaw/workspace-maya/SOUL.md |
| OLIVER's personality | ~/.openclaw/workspace-oliver/SOUL.md |
| Shared brain files | ~/.openclaw/workspace/company-brain/ |
| Provider | Dashboard | What to Check |
|---|---|---|
| DeepSeek | platform.deepseek.com | Token usage, remaining credits, API keys |
| Anthropic | console.anthropic.com | API usage, billing, rate limits |
| Ollama (local) | ollama list in Terminal | Downloaded models, status |
| Decision | What We Chose | Why |
|---|---|---|
| Primary AI Brain | DeepSeek V3 API | 95% cheaper than GPT-4o, smart enough for agents, 5M free tokens |
| Backup AI Brain | Ollama / Llama 3.1 70B | Free, offline, runs on M3 Max. Insurance policy. |
| Coding Agent Brain | Claude Sonnet 4.6 | Best coder available. Only used when BUILDER needs it. |
| Dashboard Port | Random (assigned at install) | Security — default ports are targeted by bots |
| Network Binding | 127.0.0.1 only | Only YOUR Mac can access the dashboard. No outside access. |
| Agent Safety | MAYA ↔ OLIVER cross-check | Two agents verify each other. Built-in safety net for big decisions. |
| Communication | Draft-only mode | Agents NEVER send messages directly. Always drafts for AZ's approval. |