Run a 4-agent AI company on Gemini's free tier. 105 daily tasks, $0/month.
A complete playbook for running an autonomous multi-agent business on a single LLM's free tier. Every script, timer, and intelligence pipeline — open source.
| Metric | Value |
|---|---|
| Daily automated tasks | ~105 |
| LLM requests per day | ~105 / 1,500 RPD |
| Monthly LLM cost | $0 |
| Monthly infra cost | ~$5 (Vercel hobby) |
| Agents | 4 (CEO, Social, Security, Advisor) |
| Scripts | 80+ |
| Systemd timers | 25 |
| Intelligence files | 19 |
| Self-healing layers | 3 (detect → AI diagnose → human alert) |
| Evolution modules | 7 (learning loop, dream cycle, arena, etc.) |
| GPU inference | Ollama ultralab:7b on RTX 3060 Ti |
┌─────────────────────────────────┐
│ Gemini 2.5 Flash │
│ Free Tier (1,500 RPD) │
└──────────┬──────────────────────┘
│
┌──────────▼──────────────────────┐
│ OpenClaw Gateway │
│ (rate limit + routing) │
└──────────┬──────────────────────┘
│
┌────────────┬───────┴───────┬────────────┐
▼ ▼ ▼ ▼
┌─────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│ Main │ │ Social │ │ Security │ │ Advisor │
│ (CEO) │ │ Bot │ │ Bot │ │ Bot │
└────┬────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘
│ │ │ │
└─────┬─────┴──────┬──────┴─────────────┘
│ │
┌────────▼──┐ ┌──────▼──────┐
│ Content │ │Intelligence │
│ Pipeline │ │ Pipeline │
│(24 calls) │ │ (15 calls) │
└───────────┘ └─────────────┘
Every post goes through generate → self-review → rewrite if score < 7/10.
Posts read POST-PERFORMANCE.md + COMPETITOR-INTEL.md before generating. The agent knows what worked, what flopped, and what competitors are doing — all at zero token cost.
Reply-checker monitors comments, generates context-aware responses, tracks thread depth (max 2 rounds per post), batch-limited to prevent rate limiting.
Cross-agent review before publishing. Security agent reviews marketing claims, CEO reviews technical accuracy.
Every Sunday: 3 agents propose priorities → CEO synthesizes into next week's plan. Reads all intelligence files.
blogwatcher (RSS) → hn-trending (HN API) → summarize (Jina Reader) → agent analyzes relevance. Data gathering costs 0 LLM tokens — it's pure HTTP.
| Category | Daily RPD | % of Free Tier |
|---|---|---|
| Content (4 agents × 2 posts × 2 calls) | 16 | 1.1% |
| Quality gate reviews | 8 | 0.5% |
| Engagement (4 agents × 1x/day) | 16 | 1.1% |
| Reply checker (2x/day) | 10 | 0.7% |
| Cross-agent engagement (2x/week) | 6 | 0.4% |
| Research chain (2x/day × ~4 calls) | 8 | 0.5% |
| Blog-to-social | 2 | 0.1% |
| Lead follow-up | 2 | 0.1% |
| Inquiry tracker | 0 | 0% |
| Health monitor | 0 | 0% |
| Competitor watch | 0 | 0% |
| Post stats | 0 | 0% |
| Weekly strategy | 1 | 0.07% |
| Total | ~105 | 7% |
| Remaining for interactive | ~1,395 | 93% |
The expensive part of AI agents isn't the LLM — it's wasted context.
Long conversations burn tokens on repeated context. Short, focused tasks with pre-computed data are 100x more efficient.
❌ Typical agent: "Read this RSS feed, analyze it, remember it, now write a post"
→ 1 long conversation, 10K+ tokens of context accumulation
✅ Our approach:
1. blogwatcher scan (0 LLM tokens — pure HTTP)
2. summarize URL (0 LLM tokens — Jina Reader API)
3. Read RESEARCH-NOTES.md + POST-PERFORMANCE.md (0 tokens to produce)
4. One focused prompt with all context pre-injected
5. One response → parse → post → done
05:00 Research chain → RESEARCH-NOTES.md
06:00 Customer insights sync + Inquiry tracker
06:30 Competitor watch → COMPETITOR-INTEL.md
07-10 Autopost (all 4 agents, quality-gated)
10:00 Engagement (staggered across agents)
11:00 Reply checker (conversation threading)
14:00 Blog-to-social (if new content)
17:00 Research chain (round 2)
22:00 Post stats → POST-PERFORMANCE.md
23:00 Reply checker (round 2)
Sun Weekly strategy session
free-tier-agent-fleet/
├── scripts/
│ ├── core/
│ │ ├── moltbook-autopost.sh # Content pipeline (quality-gated)
│ │ ├── team-context.sh # Cross-agent collaboration
│ │ └── peer-review.sh # Pre-publish review
│ ├── intelligence/
│ │ ├── research-chain.sh # RSS + HN → summarize → analyze
│ │ ├── competitor-watch.sh # Track competitor content
│ │ ├── post-stats.sh # Track own post performance
│ │ └── hn-trending # HN top stories (0 cost)
│ ├── engagement/
│ │ ├── reply-checker.sh # Auto-reply with threading
│ │ └── blog-to-social.sh # Auto-promote blog posts
│ ├── operations/
│ │ ├── inquiry-tracker.js # CRM pipeline tracking
│ │ ├── lead-followup.js # Auto-draft outreach
│ │ ├── health-monitor.sh # Endpoint monitoring
│ │ └── weekly-strategy.sh # AI strategy sessions
│ ├── self-heal/ # 🆕 3-tier autonomous healing
│ │ ├── self-heal.sh # Main sentinel (every 10 min)
│ │ ├── predictive-ops.sh # Anomaly trend detection
│ │ ├── tg-commander.js # Remote control via Telegram
│ │ └── fixes/ # 6 pre-written fix scripts
│ ├── evolution/ # 🆕 Self-improving intelligence
│ │ ├── apply-strategy.js # Closed-loop learning
│ │ ├── dream-cycle.sh # Idle GPU utilization
│ │ ├── agent-arena.js # Competitive evolution
│ │ ├── quality-gate.sh # Pre-publish scoring
│ │ └── memory-consolidate.sh # Weekly wisdom synthesis
│ └── content-pipeline/ # 🆕 Cross-platform distribution
│ ├── smart-post.sh # Ollama → quality gate → publish
│ └── devto-crosspost.cjs # Auto cross-post to Dev.to
├── timers/ # systemd timer + service files
├── examples/
│ ├── openclaw.json.example # Agent fleet config
│ └── credentials.json.example # API credentials
├── LICENSE # MIT
└── README.md
- Linux (WSL2 works great on Windows)
- OpenClaw installed
- Gemini API key (free tier from AI Studio, NOT from Google Cloud)
- Node.js 18+
- Optional:
blogwatcher,summarize(Jina Reader CLI)
# 1. Clone this repo
git clone https://github.com/ppcvote/free-tier-agent-fleet.git
cd free-tier-agent-fleet
# 2. Configure OpenClaw
cp examples/openclaw.json.example ~/.openclaw/openclaw.json
# Edit with your API keys and agent names
# 3. Set up credentials
mkdir -p ~/.config/moltbook
cp examples/credentials.json.example ~/.config/moltbook/credentials.json
# Edit with your Moltbook API key
# 4. Copy scripts
cp -r scripts/* ~/.openclaw/scripts/
chmod +x ~/.openclaw/scripts/**/*.sh
chmod +x ~/.openclaw/scripts/intelligence/hn-trending
# 5. Install timers
cp timers/openclaw-*.timer timers/openclaw-*.service ~/.config/systemd/user/
systemctl --user daemon-reload
# 6. Enable timers
for timer in ~/.config/systemd/user/openclaw-*.timer; do
systemctl --user enable --now "$(basename "$timer")"
done
# 7. Create workspace directories
mkdir -p ~/.openclaw/{workspace,workspace-social,workspace-security,workspace-advisor}
mkdir -p ~/.openclaw/{data,logs}
# 8. Verify
systemctl --user list-timers | grep openclaw# Test the autopost pipeline manually
bash ~/.openclaw/scripts/core/moltbook-autopost.sh
# Test the research chain
bash ~/.openclaw/scripts/intelligence/research-chain.sh
# Test health monitoring
bash ~/.openclaw/scripts/operations/health-monitor.sh- Add agent config to
openclaw.json - Create workspace:
mkdir ~/.openclaw/workspace-newagent - Copy and modify an autopost script with new pillars
- Add corresponding systemd timer
Edit the PILLARS array in moltbook-autopost.sh:
PILLARS=(
"your-niche-1"
"your-niche-2"
"your-niche-3"
"your-niche-4"
"your-niche-5"
)The system uses ~7% of the 1,500 RPD free tier. You can:
- Increase posting frequency (more posts/day)
- Add more agents
- Enable peer review on every post (currently optional)
- Run research chain 3x/day instead of 2x
Scripts are written for Moltbook but the pattern works for any platform with an API:
- Replace Moltbook API calls with your platform's API
- Keep the intelligence pipeline (it's platform-agnostic)
- Keep the quality gate (it's LLM-agnostic)
Created API key from a billing-enabled GCP project. Thinking tokens ($3.50/1M) ate everything with no rate limit cap. Fix: Always create keys from AI Studio, never from billing-enabled projects.
Pillar rotation used day_of_year % 5 — same result all day. Fix: (day * 2 + hour/12) % 5 for per-post variation.
Reply-checker processed all pending comments in one batch, triggering the platform's anti-spam. Agent account suspended for 24h. Fix: head -5 batch limit per agent per run.
Health check called getUpdates which conflicted with the gateway's long-polling. 18 duplicate messages in 3 minutes. Fix: Never call getUpdates outside the gateway.
Blogwatcher returned 30 new URLs, each triggering a summarize + analyze. 60 LLM calls in one run. Fix: head -5 cap on new URLs per run, seen-URL tracking.
| Component | Tool | Cost |
|---|---|---|
| Agent framework | OpenClaw | Free |
| LLM | Gemini 2.5 Flash (1,500 RPD) | Free |
| Runtime | WSL2 Ubuntu, systemd timers | Free |
| RSS reader | blogwatcher | Free |
| URL summarizer | Jina Reader | Free |
| HN data | HN Firebase API | Free |
| Hosting | Vercel hobby tier | ~$0-5/mo |
| Database | Firebase Firestore | Free tier |
| Notifications | Telegram Bot API | Free |
| Total | ~$0-5/mo |
3-tier autonomous healing — the fleet fixes itself before you even notice.
Every 10 min:
Layer 1 (Sentinel) → Detect 6 issue types → run pre-written fix scripts
Layer 2 (Gemini) → If fix fails → collect logs → ask Gemini for diagnosis
Layer 3 (TG Alert) → Send problem + fix attempts + AI suggestion to human
6 auto-fix scripts:
fix-ollama.sh— restart Ollama + warm up modelfix-gateway.sh— kill stale processes + restart gatewayfix-tg-polling.sh— restart for TG long-polling stallsfix-disk-space.sh— trim logs, clean temp, vacuum journalfix-failed-timers.sh— reset failed systemd timer servicesfix-gemini-ratelimit.sh— set 1-hour cooldown lock
Predictive ops — collects 6 metrics every 10 min (rolling 144 samples = 24h), detects anomaly trends before they become outages: response time creep, disk fill projection, error rate spikes, memory pressure.
TG Commander — 8 remote control commands via Telegram: /status, /restart, /leads, /post, /dream, /heal, /arena, /weights
See scripts/self-heal/ for all scripts.
The fleet doesn't just run — it gets smarter every week.
Analyzes post performance → calculates topic/language weights → writes strategy directives. The fleet automatically doubles down on what works and reduces what doesn't.
6x/day during off-peak hours, only when GPU is idle:
- Pre-generate draft posts for next day
- Competitor content analysis
- FAQ bank building from common questions
- Self-reflection and improvement notes
Weekly competitive ranking: 60% engagement metrics + 40% Ollama quality review. Top performer gets more posting slots. Underperformer gets prompt mutation suggestions.
Ollama scores every post 1-10 before publishing. Score < 6 → blocked, saved to drafts for human review. Integrated into smart-post.sh automatically.
Weekly: reads all intelligence (arena scores, strategy weights, dream reflections, heal logs, performance data) → Ollama synthesizes into WISDOM.md → copies to all agent memory directories. Persistent intelligence across weeks.
Auto cross-post blog articles to Dev.to with canonical URLs pointing back to your site. One article per day, 40-day content drip with zero manual effort.
See scripts/evolution/ and scripts/content-pipeline/ for all scripts.
PRs welcome. Particularly interested in:
- Adapters for other platforms (Twitter/X, LinkedIn, Reddit)
- Additional intelligence pipelines
- Cost optimization techniques
- Multi-LLM support (fallback chains)
MIT
Built by Ultra Lab in Taiwan. Running in production since March 2026.
Blog post: 免費仔的極限：1,500 RPD 跑 105 個自動化任務
See the agents work live: ultralab.tw/agent