Claw
The gap between wanting an agent and running one
Everyone we talk to wants the same handful of agents. Answer my Telegram overnight and send a morning digest. Triage the support inbox. A WhatsApp bot for the shop. A coding agent I can message from my phone.
None of these are hard problems anymore. The runtimes exist and they're good: OpenClaw for a full personal-agent gateway, PicoClaw when you want 16 chat channels in a 10 MB binary, Claude Code and Codex for serious coding work, Deep Agents for planning and subagents. What's hard is everything around them. Picking one. A server that stays up. Getting the Telegram token wired in without pasting it into a config file you'll forget about. Knowing, three weeks later, what the thing has been doing and what it has been spending.
That last part is where we started. ClawMetry is an open-source observability dashboard for AI agents: it exists because agents spawn sub-agents, burn tokens and call tools faster than anyone can follow. Half a million installs later, the most common question from users wasn't about dashboards at all. It was: can you just run the agent for me too?
So now we do.
Describe it. That's the whole interface.
Open build.clawmetry.com and type what you want, the way you'd explain it to a colleague. An AI engineer reads your description, asks the one or two questions that actually matter, and recommends an engine with reasons you can read. You can override it; most people don't need to.
Then it builds. Every agent gets:
- Its own micro-VM. Not a shared container pool: a dedicated, isolated VM with its own private network. Your secrets are sealed with end-to-end encryption before they leave the control plane; the infrastructure API only ever sees ciphertext.
- The right engine. 13 hostable runtimes: OpenClaw, NanoClaw, PicoClaw, Hermes, Claude Code, Codex, OpenCode, Qwen Code, Aider, Goose, Pi, Deep Agents and Cursor.
- Your channels. Telegram, WhatsApp, Slack, Discord, Signal, email, iMessage and more, depending on the engine. Plus a built-in web chat so you can talk to the agent seconds after it boots, before any channel is connected.
- A model, immediately. Start on the included Claude model with pay-per-use pricing and no setup, or bring your own Anthropic or OpenAI key. Switch either way, per agent, whenever you like.
You watch every move
Here's the part that makes this a ClawMetry product and not another agent-hosting waitlist: observability isn't a tab we added at the end. It's the reason the company exists.
Every hosted agent comes with a live activity feed, spend for today, message counts, and a fleet view across everything you're running. Every token is tied to the task that spent it. When an agent loops, you see it looping. When it's quietly perfect for a week, you see that too, and what the week cost you.
Most agent platforms treat "what is it actually doing?" as an enterprise feature. We think it's the first feature.
Grow out of an engine, not out of the platform
Agents outgrow their first engine. A chat bot picks up coding duties; a one-shot coding agent needs to become always-on. The builder handles migration: it compiles a handoff briefing from your original request and everything the old agent learned, and ships it to the new agent's memory. You keep the history, the context and the channels; only the engine changes.
Pricing that fits on a sticky note
Hosting is per agent: $7/month for a 1 GB micro-VM (most engines), $15/month for the 2 GB engines (OpenClaw, NanoClaw). Model usage is pay as you go on the included model, or free of markup on your own key. Exploring, describing and designing your agent costs nothing, and you don't need a card or an API key to start.
What do you want to build?
The builder is live today at build.clawmetry.com. Describe your agent. We build it, host it, and you watch every move.
And if your use case is bigger than a sentence, the founder-call link is inside the product. We build custom agents too.