OpenClaw Telegram Commands: The Complete Power User Guide
MCP just hit 97 million installs. Agentic AI is the new normal. But most people are still chatting with their agent like it's a search bar. Here's how to actually command OpenClaw — every slash command, every shortcut, every trick.
Why Telegram Is the Power Interface
The AI tooling world just got a reality check. Ollama shipped MLX support for Apple Silicon this week — local models are now meaningfully faster on MacBooks and Mac minis. MCP hit 97 million installs. GPT-5.4, Gemini 3.1 Ultra, Grok 4.20 are all shipping. The agentic moment is here.
And yet: most power users are still interacting with their AI agent like it's a chat window. Text in, text out. No real control surface. No execution-layer commands. Just vibes.
OpenClaw's Telegram integration is different. It's not a chat UI bolted on top — it's a command interface. You can approve dangerous operations, kill runaway agents, schedule cron jobs, toggle reasoning mode, and inspect session state — all from your phone, while on the move.
Get the weekly AI agent digest 🦞
What's shipping in AI tools, every Monday. No fluff.
🦾 The mental model shift
Stop thinking of Telegram as a chat app. Think of it as an SSH terminal in your pocket — one that connects directly to an AI agent with eyes, hands, and memory.
One-Time Setup
If you haven't connected Telegram yet, it takes under 5 minutes. Here's the fast path:
Create a bot via @BotFather
Message @BotFather on Telegram → /newbot → follow prompts → copy your token.
Add to openclaw config
Edit ~/.openclaw/config.yaml and set channel: telegram with your bot token.
Set your user ID
Message @userinfobot on Telegram to get your numeric user ID. Add it to allowedUsers in config.
Restart OpenClaw
Run: openclaw gateway restart — your bot is live.
# Minimal telegram config block
channel:
kind: telegram
token: "7812345678:AAF-your-bot-token-here"
allowedUsers:
- 123456789 # your numeric Telegram user ID
reactions: minimal # or full / offNeed the full setup walkthrough? See the complete setup guide →
Core Commands Reference
These are the commands you'll use every day. All are sent as plain messages to your bot.
/statusShow session status — model, token usage, cost estimate, reasoning mode, session key. Your quick health check.
/helpList all available commands with brief descriptions. Useful when you forget the exact syntax.
/clearClear the conversation context. Keeps memory.md intact. Use this when the model gets stuck in a loop or you want a fresh session.
/resetFull session reset. Clears context and reload all config files. Like restarting but without killing the process.
/model <name>Switch the active LLM mid-session. Example: /model anthropic/claude-opus-4 or /model openai/gpt-4o. No restart needed.
/reasoningToggle extended thinking / reasoning mode. When on, the model thinks longer before answering. More accurate, slower, higher cost.
/verboseToggle verbose mode — shows tool call details inline as the agent works. Great for debugging automations.
Agent Control Commands
These are the commands that make OpenClaw a real control surface — not just a chat window. When your agent is running autonomously, these are your levers.
# Kill a runaway sub-agent
/subagents list # → agent-abc123 (running, 4m 12s) # → agent-def456 (running, 12m 01s) ← this one is stuck /subagents kill agent-def456 # ✅ Agent killed
/subagents listShow all running sub-agents with their IDs, start times, and status. The first thing you reach for when something feels stuck.
/subagents kill <id>Terminate a specific sub-agent by its ID. Instant. The agent stops mid-execution without cleanup.
/approve allow-once <code>Approve a single elevated command. The code comes from the agent when it needs elevated shell access.
/approve allow-always <code>Grant persistent approval for a specific command pattern. Use carefully — only for trusted, well-understood commands.
/approve deny <code>Reject the pending elevated command. The agent gets a denial response and can handle it gracefully.
/sessions listList all active sessions including background agents, cron-spawned jobs, and ACP coding agents.
Managing Cron Jobs via Telegram
You don't need a YAML file to manage cron jobs. You can create, pause, and delete scheduled tasks entirely through natural language in Telegram. The agent translates your intent into a proper cron job definition.
# Schedule a recurring job — just talk to the agent
You: "Every weekday at 8am, check BTC price and
send me a 2-line brief on Telegram"
Agent: ✅ Cron job created
Name: morning-btc-brief
Schedule: 0 8 * * 1-5 (Europe/Belgrade)
Next run: Monday 8:00 AM
Job ID: cron-a3f9c2d1Once jobs exist, you can control them by saying things like:
💡 Pro tip: cron run-history
Ask the agent "Show me run history for my morning brief" — it queries the job log and summarizes success/failure rates, average runtime, and last output. Great for debugging flaky automations.
The Approval Flow Explained
This trips up a lot of new users. When OpenClaw needs to run an elevated shell command — something that writes to disk, calls a system binary, or modifies your environment — it pauses and sends you a prompt:
⚠️ Elevated command approval required Command: rm -rf /tmp/old-cache && brew upgrade node Risk: medium To proceed: /approve allow-once abc123 /approve allow-always abc123 /approve deny abc123
/approve allow-onceWhen: Use this 99% of the time.
Approves just this one invocation. Next time the same command runs, you'll get asked again. Maximum safety.
/approve allow-alwaysWhen: Only for commands you fully understand and run repeatedly.
Creates a persistent allowlist entry. The agent never asks again for this exact command. Convenient but less safe.
/approve denyWhen: When you don't recognise the command or it looks wrong.
The agent gets a denial and typically tries a safer alternative or asks for clarification.
🔒 The approval code is single-use
Each code like abc123 is tied to one specific pending command. It expires once used or denied. You can't reuse a code from a previous session — the agent always generates a fresh one.
Power Patterns & Shortcuts
These are the usage patterns that separate occasional users from people who've genuinely integrated OpenClaw into their workflow.
Pattern 1: The Model Switch Mid-Task
Switch models without restarting when a task changes
# Start with fast/cheap model for research /model openai/gpt-4o-mini "Research the top 5 DePIN projects this month" # Switch to smarter model for the actual analysis /model anthropic/claude-opus-4 "Now write a 500-word investment thesis on the top 2" # Check what you spent /status
GPT-4o-mini for data gathering, Claude Opus for reasoning. You get the best of both without paying full price for the whole session.
Pattern 2: Reasoning Toggle for Hard Problems
Extended thinking when accuracy matters more than speed
# Normal mode for everyday tasks "What's the weather in Belgrade today?" # Turn on reasoning for complex decisions /reasoning "Review this smart contract for security vulnerabilities and give me a risk assessment" # Turn it back off when done /reasoning
Reasoning mode costs more tokens but the quality jump on multi-step problems is significant. Don't leave it on permanently.
Pattern 3: Verbose Debugging
See exactly what tools the agent is calling
/verbose # turn on tool call visibility "Check my GitHub PRs and summarize what needs review" # Output now shows: # → tool: github (action: list PRs) # → tool: web_fetch (fetching PR diffs) # → Summarizing 3 open PRs... /verbose # turn off when debugging is done
Invaluable for understanding why an automation isn't doing what you expect. Always start here before assuming a bug.
Pattern 4: Inline Buttons for Yes/No Decisions
OpenClaw sends action buttons — tap instead of typing
When OpenClaw needs a binary decision (proceed / cancel, approve / deny), it renders inline Telegram buttons. You tap — no typing required. This is configured with capabilities: inlineButtons in your config.
# In openclaw config.yaml
channel:
kind: telegram
token: "your-token"
capabilities:
- inlineButtons # enables tap-to-approve UI
reactions: minimal # emoji reactions on key messagesWhat the Community Is Saying
With Ollama + MLX landing on HN today (424 points, 199 comments), local AI inference is top of mind. We pulled threads from the OpenClaw community on how people are actually using the Telegram command layer:
“The /approve flow changed how I think about AI safety. It's not a blocker — it's a checkpoint. I approve maybe 2-3 things a day and they're always the right things to pause on.”
@tomek_dev
“I use /verbose every time I'm setting up a new automation. Once it works, I turn it off. It's like having a debugger for your AI agent. Nothing else gives you this visibility.”
@ana_builds
“Switching models mid-session with /model is underrated. I run cheap models for first drafts and bring in Claude Opus for the final pass. Cost dropped 60% on my daily workflows.”
@marc_vc
“The Ollama MLX news today is huge for my setup. Running a Mac mini as an OpenClaw host — now I can run local models AND have Telegram control for remote management. That's the dream setup.”
@pita_infra
“People sleep on /reasoning. Turn it on for contract reviews, token economics analysis, anything where being wrong is expensive. The quality difference is night and day.”
@dex_protocol
Start Commanding Your Agent
The difference between a chatbot and an agent is control. A chatbot waits for you. An agent acts — and you need real commands to guide it, correct it, and push it further. That's what the Telegram layer gives you.
# Your first 5 minutes with OpenClaw on Telegram
# 1. Check your session /status # 2. Try the agent "What's the weather in Belgrade and top tech news today?" # 3. Turn on verbose to see how it works /verbose "Do the same thing again" /verbose # 4. Create your first automation "Every morning at 8am, send me BTC price + top 3 crypto news headlines on Telegram" # 5. List your cron jobs "Show me all scheduled jobs" # You're now running a local AI agent. # Nobody else has access to this infrastructure. # This is yours.
Get the AI Adaptation Playbook
12 pages. 5 frameworks. 6 copy-paste workflows. Everything you need to future-proof your career with AI.
Instant delivery · No spam · Unsubscribe anytime