OpenClaw Intro: Building a Twitter Bot with Raspberry Pi and Gemini

Setup guide — Raspberry Pi as a sandboxed AI agent environment

Motivation

"I want AI to pull the latest baseball news and tweet it automatically."

That's how this project started. I built an automated baseball tweet bot using Raspberry Pi 5 + OpenClaw + Gemini API (free tier). Running cost: ~$2/month (electricity only).

Most AI bot tutorials you'll find use:

• ChatGPT API → pay-per-use (at least $5/month)
• Claude API → pay-per-use
• AWS Lambda + DynamoDB → charges when you exceed the free tier
• Heroku / Railway → free plans discontinued or heavily restricted

This post is for people who want to experiment cheaply. If you have a Raspberry Pi sitting around, you can start today.

What It Produces

The bot posts tweets like this 13 times a day, fully automatically:

WBC 2026 starts March 5th! Can't wait! The excitement is like the silence before the Earth's core starts boiling. ⚾

The character is "a slightly quirky baseball-obsessed old guy." The tweet quality is, well, what it is — but it runs.

Architecture & Cost

[cron schedule] → [OpenClaw Gateway] → [Gemini 2.5 Flash API]
                                              ↓
                                        web_search for latest news
                                              ↓
                                        write tweet (140 chars)
                                              ↓
                                    [tweet.js] → Twitter API

Monthly running cost: ~$2 (electricity only)

Why Raspberry Pi?

"Why not just use a VPS?" Fair question. Here's why I went with Pi:

1. Perfect sandbox — When you let AI run shell commands, you want it completely isolated from your main PC. Worst case, you re-flash the SD card
2. Always-on by design — VPS free tiers have time limits or get suspended. Pi just stays on
3. Low learning curve — SSH in, run commands. No Kubernetes, Docker, or AWS needed
4. One-time cost — No monthly billing, so you can leave it running without worry

Tech Stack

What is OpenClaw?

OpenClaw is an open-source AI agent framework. In short:

• Gateway for LLMs (Gemini, Claude, Ollama, etc.)
• Runs scheduled tasks via cron
• Has an exec tool — AI can run shell commands directly
• Has web_search for real-time information (a Brave Search API key may be required)
• Integrates with Discord, Slack, Telegram, etc.

The flow here is: cron → Gemini → web_search → exec (tweet.js)

Gemini 2.5 Flash Free Tier

Get an API key from Google AI Studio and use Gemini 2.5 Flash for free:

• 15 RPM (requests per minute)
• 1500 RPD (requests per day)

13 cron jobs/day is well within the limit. I space jobs at least 1 hour apart.

Twitter API Free Tier

500 posts/month limit. At 13 tweets/day × 30 days = 390 posts — fits comfortably. Write-only (no timeline read), but that's fine for a bot.

Setup (Overview)

Full instructions are in the repository README. Here's the high-level flow.

1. Install OpenClaw on Pi

curl -fsSL https://openclaw.ai/install.sh | bash
openclaw onboard  # enter your Gemini API key

2. Add the tweet script

tweet.js is a simple ~30-line Node.js script:

// simplified
const { TwitterApi } = require("twitter-api-v2");
const client = new TwitterApi({ /* OAuth 1.0a keys */ });
const text = process.argv.slice(2).join(" ");
await client.v2.tweet(text);

3. Define the bot's personality (SOUL.md)

OpenClaw reads SOUL.md from the workspace to set the agent's character:

# Baseball Guy

## Tweet style
- Casual, like muttering at a bar
- Under 140 characters
- RULE: include one grandiose, off-the-wall metaphor every time

4. Register cron jobs

openclaw cron add \
  --cron "0 9 * * *" \
  --tz "Asia/Tokyo" \
  --name "WBC 09:00" \
  --system-event "Look up the latest WBC news and tweet in under 140 characters."

Every day at 9am: Gemini wakes up, searches the web, writes a tweet, posts it. Fully automatic.

5. Persist with systemd

systemctl --user enable openclaw-gateway
systemctl --user start openclaw-gateway

Auto-starts when Pi boots. No need to SSH in.

The Ollama Rabbit Hole (Bonus)

I originally tried running everything offline with Ollama (local LLM). No API limits, full privacy — sounds ideal, right?

It didn't work on Raspberry Pi 5 + llama3.2:3b.

The traps I fell into

1. OpenClaw hardcodes contextWindow at 128K → Instant OOM on Pi. Patched dist files to 8K
2. CONTEXT_WINDOW_HARD_MIN is 16K → Conflicts with 8K context. Another dist patch
3. Tools schema alone is 17K characters → Disabled most tools via tools.deny
4. OLLAMA_LOAD_TIMEOUT=0 means "default 5 min", not "unlimited" → Subtle trap
5. stream:true causes GIN server to timeout after 5 min of silence → Patched to stream:false
6. prompt eval: 4.1 tokens/sec → 3800-token prompt takes 15 min. HTTP 500 at 5 min
7. Swap was only 200MB → Expanded to 2GB

After solving all of that, llama3.2:3b's tool calling accuracy was terrible:

• Passes "{}" (string) to env argument instead of an object → validation error
• Outputs tool calls as plain text instead of function calls
• When asked about the Hanshin Tigers, replied in English: "Let's go Orix!" (wrong team entirely)

After 2 days of fighting, I switched to Gemini. It posted a tweet in 18 seconds. Sometimes you just have to let go of what you're attached to.

See OLLAMA_TROUBLESHOOTING.md for the full breakdown.

Security

Letting AI run shell commands requires care. Using Pi as a sandboxed environment keeps the risk contained:

• Dedicated machine — no personal data from main PC
• UFW firewall — block all incoming except SSH
• Gateway bound to localhost — not reachable externally
• tools.deny — AI can only use exec and web_search
• .env permissions 600 — API keys locked down

If the AI goes rogue, the damage stays on the Pi. Worst case: re-flash the SD card.

Summary

Before spending money on cloud infrastructure, try Raspberry Pi first. It's disposable, always-on, and free to run after the initial hardware cost.

Source code: github.com/yasumorishima/raspi-baseball-bot