By the end of this guide you will have a personal AI agent that you can message from Telegram or WhatsApp, running on your own computer or server, with its own memory, limits on what it may do, and a routine for keeping it safe. The job is done by an open-source agent runtime (or a no-code agent service): you install it, connect a model provider and a chat channel, and then talk to it like a contact in your messaging app.
If you want to run a personal AI agent on Telegram or WhatsApp, the main choices are where it runs, which model pays for the thinking, and how much the agent may do without asking. This guide covers each choice and compares seven tools from our Chat assistants category.
Key takeaways
- Pick a runtime, install it on a machine you control, add a model provider key and a chat channel, then test with harmless requests before you give it any real access.
- The main decision is how much you want to run yourself: a self-hosted open-source agent, or a no-code agent service on Telegram.
- The software is often free, but the model provider bills separately. OpenClaw, ZeroClaw, nanobot and QwenPaw list no subscription price, and Hermes Agent is free under the MIT license with an optional Nous Portal subscription.
- An agent that can run commands and read files is powerful and risky. Keep approvals on, limit its access, and never leave an exposed agent open to strangers.
What you need before you start
- A machine that stays on. A laptop, a spare desktop, a home server or a cloud server. A local agent does not run scheduled tasks while it is stopped.
- A model provider. An API key from a cloud provider, or a local model runtime. Model costs are usually separate from the agent software.
- A messaging account. A Telegram bot token for Telegram, or a WhatsApp connection where the tool supports it.
- Comfort with a terminal. Most of these tools are installed from the command line, although some have a desktop app or a browser setup page.
- Time. A first working agent takes about an hour or two; tuning permissions takes longer.
- Difficulty. Moderate for self-hosted tools, easier for the no-code service.
- Consent. Do not connect other people's chats or contacts to an agent without their agreement.
Which kind of AI tool fits this job
Choose by how much you want to manage yourself.
Full-featured open-source agents. Hermes Agent from Nous Research is free under the MIT license, keeps persistent memory, runs scheduled tasks in plain language, and is reachable from messaging platforms including Telegram and WhatsApp. OpenClaw runs on your own devices and connects through a single gateway to Telegram, WhatsApp, Discord, Slack and other chat apps, although its core install ships only A2A, Reef, Telegram and WebChat, and other channels are plugins. Its listing says there is no subscription or hosted tier.
Safety-first or lightweight runtimes. ZeroClaw is a single Rust binary that connects many model providers to many chat channels, including Telegram, and defaults to supervised autonomy: medium-risk actions wait for approval and high-risk ones are refused. Mercury Agent asks before it writes files or runs shell commands, and can be reached from the terminal, Telegram, Discord, Slack or Signal. nanobot is a lightweight agent with a browser setup page and chat channels including Telegram.
Agents with their own memory system. QwenPaw keeps a three-layer memory and offers a sandbox and guards that block dangerous commands. It reaches chat channels including Telegram and can run local models without an API key.
A no-code route. Mangaba AI offers Mangaba Agent, a four-step flow that puts an agent on Telegram without coding: create an account, connect your own AI provider key, paste a Telegram bot token and activate. Its product pages are in Portuguese.
Step 1: Decide what the agent is for
A narrow first job keeps the risk small.
- Pick two or three tasks, such as answering questions, summarizing notes or running a daily reminder.
- Decide what the agent must never do, such as send email or delete files.
- Choose the channel. Telegram has the widest support across these tools, while WhatsApp support is listed for Hermes Agent and OpenClaw, and appears among ZeroClaw's integrations.
- Decide where it will run, and who else will be allowed to message it.
Check it worked by writing a short list of allowed tasks and a list of forbidden ones.
Step 2: Choose a runtime and a model provider
You end this step with a tool and a way to pay for the model.
- For the most features, look at Hermes Agent or OpenClaw. For tight approvals, look at ZeroClaw or Mercury Agent. For a no-code start on Telegram with a free plan, look at Mangaba AI.
- Pick a model provider. Cloud providers need an API key. QwenPaw and others can run local models through tools such as Ollama or LM Studio instead.
- Check what is free and what is billed separately in the comparison table below.
- Choose a machine that matches the tool. ZeroClaw lists Linux, macOS, Windows and Raspberry Pi, and Hermes Agent lists macOS, Windows and Linux.
Check it worked by confirming your machine and provider are supported by the tool you chose.
Step 3: Install the agent
Install on a machine you control and keep it separate from sensitive files if you can.
- Follow the vendor's install instructions. Depending on the tool this is a script, Homebrew, Docker, a desktop app or a Python package. QwenPaw's pip install needs Python 3.11 or newer.
- Create a dedicated user account or a container for it where possible, so the agent cannot see your whole disk.
- Start the agent and open its setup page or terminal interface.
- Add your model provider key through the setup flow, not in a shared file.
Check it worked by sending a first message in the tool's own interface and getting a reply.
Step 4: Connect Telegram or WhatsApp
Connect one channel and keep access closed to everyone but you.
- For Telegram, create a bot and paste its token into the agent's channel settings. Mangaba Agent's setup asks for exactly this.
- For WhatsApp, follow the tool's channel setup. Because WhatsApp is not a simple bot channel in every tool, read its documentation closely.
- Limit who can message the agent. OpenClaw, for example, requires unknown senders to be approved through DM pairing before they get replies.
- Keep tokens and keys out of screenshots and shared notes.
Check it worked by messaging the agent from your phone and getting a reply, then confirming a second account is blocked.
Step 5: Set permissions and approvals
An agent that can act needs boundaries.
- Turn on approval prompts for writes, shell commands and sending messages. ZeroClaw defaults to supervised autonomy, and Mercury Agent waits for approval on writes, shell commands and pushes.
- Use sandboxing where offered. ZeroClaw lists Landlock, Bubblewrap, Seatbelt and Docker. Hermes Agent can run code in local, Docker, SSH, Singularity or Modal backends.
- Give it the smallest set of tools and accounts it needs.
- Test a risky request, such as deleting a file, and confirm it asks first or refuses.
Check it worked by seeing the agent stop and ask before anything risky.
Step 6: Teach it memory, skills and routines
Once it is safe, make it useful.
- Tell it your preferences and the context for your tasks, and check what it stored. Several tools keep persistent memory, and Mangaba Agent lets you read and delete what the agent remembers.
- Add skills for repeated jobs. Hermes Agent can write reusable skills from experience, and QwenPaw and others list skills.
- Set a scheduled task, such as a daily summary, and confirm the machine will be on when it runs.
- Review what it did after a few days, and tighten anything it should not have been able to do.
Check it worked by receiving a scheduled message on time, with content you expected.
Step 7: Maintain it
A running agent needs the same care as any small server.
- Keep the software updated. nanobot's listing warns that releases change quickly, so back up config and sessions before upgrading.
- Watch model usage, since the provider bills separately.
- Rotate keys if a device is lost, and revoke access you no longer need.
- Check your logs from time to time for requests you did not make.
Check it worked by being able to list what the agent can access and what it cost last month.
Tools for this job compared
| Tool | Good fit for | Pricing | Stated limits |
|---|---|---|---|
| Hermes Agent | Persistent memory and scheduling, reachable from Telegram, WhatsApp and more | Open source | Model providers charge separately; tasks need an active agent, so a stopped local agent does not run them |
| OpenClaw | One gateway to Telegram, WhatsApp, Discord, Slack and others | Open source | Only A2A, Reef, Telegram and WebChat ship in the core install; you run the gateway yourself |
| ZeroClaw | Approval-gated, sandboxed agent on your own hardware | Open source | No hosted version or account system; you maintain it; project is on the v0.8 series |
| nanobot | Lightweight agent with a browser setup page | Open source | You pay your own model provider; releases change quickly, so back up before upgrading |
| QwenPaw | Layered memory and local models, including Telegram | Open source | Desktop application is in beta; cloud providers need an API key; pip install needs Python 3.11 or newer |
| Mangaba AI | No-code agent on Telegram with your own provider key | Freemium, paid from R$39/mo (checked 24 Oct 2026) | Free plan includes 1 agent, Pro allows up to 3 and Empresa up to 10; product pages are in Portuguese |
| Mercury Agent | Permission-gated agent for terminal and chat apps | Open source, paid from $19/mo (checked 24 Oct 2026) | Mercury Cloud Free includes 1 agent connection and 5,000 Mercury Credits per month |
Prices and plan limits change, so confirm the current terms on each vendor's site before you subscribe.
Common mistakes to avoid
- Giving the agent broad access on day one. An agent that can run commands can also run the wrong ones. Fix it by starting with approvals on, a sandbox and a minimal set of tools.
- Leaving the bot open to strangers. Anyone who finds a bot can message it. Fix it by restricting who can talk to it, and by requiring approval for new senders, as OpenClaw does with DM pairing.
- Forgetting model costs. The software may be free while the provider bills by use. Fix it by setting a budget or limit with the provider and checking usage regularly.
- Storing secrets carelessly. Tokens and keys leak through screenshots and shared files. Fix it by keeping them in the tool's settings and rotating them if exposed.
- Connecting other people's chats without asking. Messages from friends and colleagues may be personal. Fix it by getting agreement first, and by keeping the agent out of group chats where it is not wanted.
How we put this together
This is an evidence-based guide built from each tool's listing in our directory, which is compiled from the vendor's own site and pricing page. It is not a review based on using the tools. We don't rank tools, and nobody paid to be included. Some outbound links may be affiliate links; see our disclosure.
Frequently asked questions
Is there a free way to run a personal AI agent on Telegram?
Yes, for the software. OpenClaw, ZeroClaw, nanobot and QwenPaw list no subscription price, and Hermes Agent is free under the MIT license, with Nous Portal as an optional subscription. You still pay your model provider unless you run a local model, and you need a machine to run the agent on. Mangaba Agent has a Free plan with 1 agent.
Can I use WhatsApp instead of Telegram?
Some tools support it. Hermes Agent and OpenClaw list WhatsApp, and WhatsApp appears among ZeroClaw's integrations. Support differs by tool, so read the channel documentation before you commit.
Is it safe to give an AI agent access to my files and accounts?
Only with care. Use approval prompts, a sandbox, a dedicated account and the smallest set of permissions. Test risky requests before relying on the agent, and keep a human decision in the loop for anything that sends messages or deletes data.
Can I keep my data private?
Partly. The agent software runs on your machine, but model calls go to whichever provider you choose unless you use a local model. QwenPaw can run local models, and Hermes Agent can use any model provider you choose, so you decide where data goes.
How long does it take to set up?
The vendors do not publish setup times, and it depends on the tool and your comfort with a terminal. Setting sensible permissions and tuning memory and routines take longer, and are worth the time.
Compare more options in our Chat assistants category.
