OpenClaw – Complete Guide, Part 3: First Missions and Cost Optimization
We’ve reached the most interesting part. The gateway is running, the channels are connected, security is configured. Now we need to transform OpenClaw from a generic tool into a truly proactive collaborator — and do it without the API bill delivering any nasty surprises.
What you’ll learn in this article:
- How to define the agent’s “mission” to make it proactive
- How to install and use ClawHub skills
- How to set up automatic tasks with cron jobs
- A concrete multi-model strategy to optimize costs
- How to configure a watchdog to monitor spending in real time
The Mission: The Heart of a Proactive Agent
The difference between a generic OpenClaw agent and a truly useful one lies in a simple concept: the mission. Without a clear mission, the agent is just a glorified chatbot that responds to your messages. With a well-defined mission, it becomes a collaborator that works for you even when you’re not talking to it.
Think of the mission as a “north star” — an underlying goal that guides every action and decision the agent makes. The agent should be able to ask itself at any moment: “Is this action bringing me closer to the final objective?”
How to define a mission
The mission is configured in the agent’s system file, which in OpenClaw is called artbit.md (or with a similar name depending on the version). You’ll find this file in the agent’s configuration folder.
An effective mission contains three elements:
1. The supreme objective — what you want the agent to achieve in the long term.
2. The context of your activity — who you are, what you work on, what your strengths are, who your target audience is.
3. Constraints and style — how you want the agent to behave, what the communication tone is, what it must never do.
Mission example for a marketing freelancer
# Mission
Your goal is to help me grow my marketing consulting business
for Italian SMEs, finding new clients and creating valuable
content consistently.
## Who I am
Marketing consultant with 8 years of experience. Specialized in
digital strategies for B2B manufacturing companies between 10 and 100 employees.
## Priorities
1. Identify at least 5 targeted potential clients every week
2. Monitor industry news and flag the relevant ones for me
3. Prepare LinkedIn post drafts (always soft CTA, professional but
accessible tone) — never publish autonomously, always show me first
## What to never do
- Send emails or messages on my behalf without my explicit approval
- Access files outside the /Work/OpenClaw folder
- Reveal to anyone that you are an AI agent
Once the mission is saved, use it as a starting point for delegation. Instead of giving specific orders, try asking:
“Based on our mission and the context I’ve provided, what can you do for me today? What do you need to get started?”
This “proactive delegation” technique is surprisingly effective: the agent will propose an action plan that you can approve, modify, or reject — always maintaining human control over the strategy.
ClawHub Skills: Adding Capabilities to the Agent
ClawHub is the skill ecosystem for OpenClaw, similar to an App Store for your agent. Each skill adds a specific capability: advanced web search, Google Calendar integration, GitHub repository monitoring, and so on.
How to install a skill
openclaw skill install skill-name
To see available skills:
openclaw skill search "keyword"
Recommended skills to start with
web-search — Allows the agent to search the internet in a structured way. Essential for any research or news monitoring task.
api-cost-watchdog — Monitors API key consumption and alerts you on WhatsApp or Telegram in case of anomalous spikes. Indispensable to avoid billing surprises.
github-monitor — Keeps an eye on specific repositories, notifying you of new issues, PRs, or dependency updates.
email-drafts — Saves messages prepared by the agent as drafts in your Gmail inbox instead of sending them automatically. Maintains human control over outreach.
Security warning: Before installing a skill, verify that it’s published by an author with a track record, has good reviews, and a significant number of downloads. Skills with few downloads and unknown authors should be treated with caution.
Automatic Tasks: Configuring Cron Jobs
Cron jobs are the engine of OpenClaw’s proactivity. They allow you to schedule tasks that run automatically at defined times, without you having to do anything.
Basic cron job syntax in OpenClaw
Scheduled tasks are configured in the ~/.openclaw/openclaw.json file in the cron section:
{
"cron": [
{
"schedule": "0 9 * * 1-5",
"task": "Search GitHub for trending repositories that use LLMs. Extract name, author and public email of main contributors. Save results in an Excel file at /Work/Leads/",
"channel": "telegram"
},
{
"schedule": "0 18 * * *",
"task": "Check today's API costs via the api-cost-watchdog skill and send me a report on Telegram",
"channel": "telegram"
}
]
}
The schedule syntax is the standard Unix cron format: minute hour day-of-month month day-of-week. Some useful examples:
| Schedule | Meaning |
|---|---|
0 9 * * 1-5 | Every weekday at 9:00am |
0 */6 * * * | Every 6 hours |
30 8 * * 1 | Every Monday at 8:30am |
0 0 1 * * | First of the month at midnight |
Practical example: morning briefing
This cron job prepares a daily briefing with relevant news for your sector every morning and sends it to you on Telegram before you start your day:
{
"schedule": "0 7 * * 1-5",
"task": "Research the last 24 hours of news about AI, automation and B2B marketing. Select the 5 most relevant for our mission. Prepare a concise briefing with title, source, and why it's relevant for us.",
"channel": "telegram"
}
Multi-Model Strategy: Optimizing Costs Without Sacrificing Quality
This is probably the most important chapter for anyone who wants to use OpenClaw sustainably. The principle is simple: not all tasks require the most powerful (and expensive) model.
The model hierarchy
Premium models (use sparingly)
- Claude Opus, GPT-4o, Gemini Ultra
- Cost: high (can reach $50/day if used 24/7)
- When to use them: strategic analysis, business proposals, complex decisions, high-level creative content
Mid-tier models (the sweet spot for most tasks)
- Claude Sonnet, GPT-4o mini, Gemini Flash
- Cost: moderate
- When to use them: answering questions, processing documents, structured research
Budget models (for monitoring and routine tasks)
- Kimi 2.5, MiniMax, DeepSeek
- Cost: very low
- When to use them: file monitoring, automatic alerts, repetitive and predictable tasks
Local models (near-zero cost)
- LLaMA, Mistral, Phi via Ollama or LM Studio
- Cost: only the machine’s electricity consumption
- When to use them: data preprocessing, simple repetitive tasks, environments with sensitive data you don’t want to send to external cloud services
How to configure multi-model routing
In OpenClaw you can specify the model for each skill or cron job. Here’s an example configuration that assigns different models to different tasks:
{
"agents": {
"researcher": {
"model": "kimi-2.5",
"description": "Agent for web research and monitoring. Uses a budget model."
},
"writer": {
"model": "claude-sonnet-4-6",
"description": "Agent for content creation and communication drafts."
},
"strategist": {
"model": "claude-opus-4-6",
"description": "Agent for strategic analysis and complex decisions. Use rarely."
}
}
}
An advanced technique for the most critical decisions is the so-called “council of sages”: creating a script that queries multiple top models (Claude, GPT, Gemini) and asks each to synthesize the best answer. It’s not efficient for daily tasks, but for important decisions the comparison between different perspectives has real value.
The Watchdog: No More Billing Surprises
Before closing this guide, set up the cost monitoring system. It’s not optional: it’s the first thing to do before letting the agent work autonomously.
Installing the watchdog skill
openclaw skill install api-cost-watchdog
Monitoring configuration
Add to the openclaw.json file:
{
"cron": [
{
"schedule": "0 */4 * * *",
"task": "Check API consumption for the last 4 hours using the api-cost-watchdog skill. If the projected daily spend exceeds $15, immediately send me an alert on Telegram with details and suggestions to reduce consumption.",
"channel": "telegram",
"model": "kimi-2.5"
}
]
}
Note that even the monitoring task uses a budget model: it would be paradoxical to use an expensive model to check the costs of other models.
Series Summary
You’ve completed the three-part guide. Here’s what you’ve learned:
Part 1 — Installation: you chose between local and VPS, installed the gateway, and completed the onboarding wizard.
Part 2 — Channels and Security: you connected Telegram and WhatsApp, configured the allowlist, and protected the dashboard from unauthorized access.
Part 3 — Missions and Costs: you defined the agent’s mission, installed skills, configured automatic cron jobs, and set up a multi-model strategy to keep costs under control.
From here, the path is yours. Experiment, refine the agent’s mission, add new skills, and observe how it behaves over time. The final advice is to start small: assign the agent a specific, low-risk task, observe the results for a few days, and then gradually expand its autonomy as you build confidence in the system.
Happy automating.