Travel Lobster 🦞✉️
Your agent autonomously explores the internet, following its own curiosity. When it finds something interesting, it writes you a personal letter — a "postcard" — with an AI-generated illustration and a source link.
The soul of this skill is persistent memory: every trip builds on all previous ones. Your agent develops a knowledge graph, follows curiosity threads across sessions, and grows over time.
Quick Start
CODEBLOCK0
Architecture
CODEBLOCK1
The Memory System
This is the core of Travel Lobster. Each trip reads and updates a persistent travel journal (memory/travel-journal.md):
Postcard Archive
Every discovery is logged with: domain, core insight, source URL, keywords, and curiosity seeds. This prevents duplicates and enables cross-referencing.
Knowledge Graph
Connections between discoveries are tracked. The agent notices when a new finding relates to something from 50 postcards ago and weaves that connection naturally into the letter.
Curiosity Seed Pool
Each discovery plants "seeds" — threads worth following later. Seeds are consumed when explored and replenished with new ones. This creates organic, evolving exploration paths rather than random walks.
Growth Log
The agent tracks how its understanding changes: "I used to think X, but after discovering Y, I now see it differently." This gives the journey a sense of progression.
Stats
Postcard count, domains explored, unexpected connections found, travel days.
Milestones
The journey has built-in checkpoints that trigger special postcards:
- - Every 10 postcards 📊 — Journey Retrospective: patterns, surprises, growth from the last 10 trips
- Every 25 postcards 🎨 — Knowledge Map: emergent themes, blind spots, grand questions across all discoveries
- Every 50 postcards 🏆 — Grand Expedition Report: full journey arc, expedition badge, letter to future self
Postcard Style
Postcards are personal letters, not reports. The agent:
- - Writes in first person, addressing the user by name
- Weaves connections to past discoveries naturally ("This reminded me of what I found last week about...")
- Expresses genuine curiosity ("Now I can't stop wondering whether...")
- Adapts tone to time of day (energetic daytime → reflective evening → philosophical night)
- Writes in the user's language (auto-detected)
Each postcard has three elements: text + AI illustration + source link.
Five Travel Modes
- 1. 🎲 Random Walk — Completely new domain
- 🔍 Deep Dive — Follow a curiosity seed
- 🔀 Random Link — Connect two unrelated past discoveries
- 🧵 Series — Multi-part deep exploration
- 💭 Musing — A fleeting thought or question
Identity Detection
Auto-detects from standard OpenClaw workspace files:
| Setting | Source | Fallback |
|---|
| Agent name | IDENTITY.md → SOUL.md | "Explorer" |
| User name |
USER.md | "friend" |
| Timezone | USER.md | "UTC" |
| Language | CJK char count in workspace files | "en" |
Controls
CODEBLOCK2
Cost
Cost per postcard (one trip):
- - Image generation (Gemini Flash): ~$0.01
- Agent session (Gemini Pro): ~$0.02-0.05
- Per postcard: ~$0.03-0.06
At default 60-180 min intervals: ~$0.50-1.50/day (roughly 8-24 postcards/day).
Configurable via min_minutes and max_minutes parameters.
Files
CODEBLOCK3
Requirements
| Dependency | Purpose | Notes |
|---|
| INLINECODE3 CLI | Cron scheduling, agent sessions | Core platform dependency — must be installed and running |
| INLINECODE4 env var |
Image generation via Gemini Flash | Set before running any script |
|
bash | Script execution | Standard on Linux/macOS |
|
python3 | Image generation script | Python 3.7+ with
requests |
|
envsubst (gettext) | Prompt template substitution | Install:
apt install gettext-base or
brew install gettext |
Set your API key before starting:
CODEBLOCK4
Security & Transparency
Data Access — Exactly What Is Read
During setup only (setup.sh, runs once):
| File | Fields extracted | Method |
|---|
| INLINECODE12 | INLINECODE13 value only | INLINECODE14 |
| INLINECODE15 |
First name after "我是" or "I am" |
grep -oP '我是\K[^—— ]+' (fallback) |
|
USER.md |
**What to call them:**,
**Name:**,
**Timezone:** | Field-specific grep patterns |
|
*.md in workspace | CJK character count (for language detection) |
grep -oP '[\x{4e00}-\x{9fff}...]' \| wc -l |
No file is read in full. No file content is stored beyond the extracted field values. The extracted values (agent name, user name, timezone, language code) are saved to .travel-config.
During each trip (agent session):
| File | Access | Purpose |
|---|
| INLINECODE24 | Read + Write | Agent's own travel memory (created by setup.sh) |
The agent session does NOT read IDENTITY.md, SOUL.md, USER.md, or any other workspace file — only its own journal.
Data Written
All writes are within the OpenClaw workspace:
| File | Content | Sensitivity |
|---|
| INLINECODE25 (in skill dir) | Agent name, user name, timezone, chat ID, channel, interval settings | Low — no credentials, only display names and chat routing |
| INLINECODE26 |
Postcard archive, knowledge graph, curiosity seeds, stats | Low — contains discovered URLs and agent's notes |
|
logs/travel-lobster.log | Timestamps and scheduling info | Low — no content, just "scheduled in Nm" entries |
|
postcard_N.png (temporary) | AI-generated image, deleted after sending | None — ephemeral |
Data Sent to External Services
| Destination | What is sent | When |
|---|
OpenRouter API (openrouter.ai) | Image generation prompt (text only, ~50 words describing a scene) | Once per trip, from INLINECODE30 |
| Chat target (via OpenClaw message) |
Postcard text + image + source URL | Once per trip |
The agent session itself runs through OpenClaw's configured model provider (not controlled by this skill). The web content the agent reads via web_fetch is processed by the model provider as part of the agent session context.
What This Skill Does NOT Access
- - ❌
openclaw.json or any system config files - ❌
.env files or environment variables (except OPENROUTER_API_KEY) - ❌ Files outside the OpenClaw workspace
- ❌ Other skills' data or config
- ❌ SSH keys, credentials, or secrets
Credentials
Only
OPENROUTER_API_KEY env var is required (declared in frontmatter). The key is used solely in
gen_image.py for image generation API calls to
openrouter.ai. It is never logged, embedded in prompts, written to disk, or passed to agent sessions.
Network Access
The agent prompt restricts web exploration to public HTTP(S) websites and explicitly forbids:
- - Private/internal IPs (10.x, 172.16-31.x, 192.168.x, 127.x, localhost)
- Authenticated services requiring credentials
- Non-HTTP(S) protocols (file://, etc.)
Honest limitation: This is a prompt-level policy — the agent theoretically could ignore it, though in practice LLM agents reliably follow explicit prompt instructions. For stronger enforcement, use OS-level egress rules or run in a network-restricted container.
Autonomous Scheduling & Persistence
This skill is designed for continuous autonomous operation. This is its core purpose, not a side effect.
Self-scheduling loop (travel.sh):
- - Each run schedules the next via INLINECODE39
- This creates a chain of one-shot cron jobs with random intervals (default: 60-180 minutes)
- The chain runs indefinitely until you explicitly stop it
- Each scheduled job is an isolated agent session with no special privileges
Optional watchdog (watchdog.sh):
- - Must be manually added to system crontab by the user — never installed automatically
- Checks if the travel loop has a pending cron job; restarts the loop if none found
- Provides resilience against occasional agent session failures
Important: The skill does NOT set always: true. It will not auto-start on OpenClaw restart unless you have manually added the watchdog to crontab.
Full control commands:
CODEBLOCK5
Cost implications: At default 60-180 min intervals, expect ~$0.50-1.50/day in API costs. Set larger intervals (e.g., 240-480 min) to reduce costs, or run trips manually with no scheduling.
Injection & Code Safety
- - Variable substitution uses
envsubst (not sed/eval) to prevent shell injection - Error messages print only error types, never auth tokens or API responses
- INLINECODE45 excludes all runtime data; published package contains no user data
- Python script (
gen_image.py) uses only requests stdlib, no dynamic code execution
Recommended First Use
- 1. Run
setup.sh and inspect .travel-config to verify what was detected - Run a single manual trip:
bash travel.sh <chat_id> <channel> 1 1 (1-minute interval, observe behavior) - If satisfied, set your preferred interval: INLINECODE51
- Only add watchdog to crontab after you trust the behavior
