ClawMetry Agent Builder
Guides
ClawMetry Agent Builder
Coding agents you can run always-on in an isolated micro-VM and message from Telegram. What each is good at, which key it needs, and what it costs.
A coding agent on your laptop stops when you close the lid. A hosted one does not — it keeps a workspace, holds context between messages you send hours apart, and answers when you message it from a phone. That is a different tool with the same name, and it is what all 9 runtimes below are.
They share a shape: a full agent with files, a shell and git, running in its own Firecracker micro-VM with a persistent volume, reachable from Telegram and every other channel in the table's runtimes. They differ in the harness, the credential you have to bring, and the temperament. This page is the comparison; each name links to its own setup guide.
| Runtime | Pick it when | Bring | Hosting |
|---|---|---|---|
| Claude Code | The reference implementation. If you have no strong opinion about which coding agent to host, host this one — it has the deepest tool use, and it is the only runtime here that can run on a Claude subscription instead of metered API tokens. | your API key, or the managed model, or a subscription | $9/mo · 1 vCPU / 1 GB |
| Codex | Pick this when your team is already standardised on OpenAI models and you want the coding agent to match. Same shape as hosted Claude Code — persistent workspace, Telegram front door — with OpenAI's harness behind it. | your API key, or the managed model, or a subscription | $9/mo · 1 vCPU / 1 GB |
| opencode | The open-source choice when you want to read the harness you are trusting with a shell. opencode is SST's agent, it is model-agnostic, and nothing about it is a black box. | your API key | $9/mo · 1 vCPU / 1 GB |
| Qwen Code | The one to host if you are running Qwen or any OpenAI-compatible endpoint — including a self-hosted or regional provider that the other harnesses will not talk to. | your API key, or the managed model | $9/mo · 1 vCPU / 1 GB |
| Aider | The only runtime here whose unit of work is a git commit. Aider edits a repo and commits, so the agent's output is a diff you can review and revert rather than a chat log you have to act on yourself. | your API key | $9/mo · 1 vCPU / 1 GB |
| goose | Choose goose when the job needs tools that are not files and a shell. Its extension system is the point — it is the most natural fit for wiring an agent into other systems. | your API key | $9/mo · 1 vCPU / 1 GB |
| pi | The lean option. pi does the coding-agent loop without a heavy harness around it, which makes it noticeably faster to first token on short tasks — the ones you actually send from a phone. | or the managed model, your API key | $9/mo · 1 vCPU / 1 GB |
| deepagents | The one that plans before it acts. deepagents runs LangChain's deep-agent loop — an explicit plan, subagents, skills — which is what you want for a task measured in hours rather than a question measured in seconds. | or the managed model, your API key | $9/mo · 1 vCPU / 1 GB |
| Cursor Agent | Host this if your team already lives in Cursor and you want the same agent reachable when nobody has the editor open. | your API key | $9/mo · 1 vCPU / 1 GB |
Each agent gets a persistent volume. Files survive restarts and idle stops,
which is what makes "clone the repo once, then message it for a month"
possible. Without that you have a chatbot that can run ls.
The micro-VM stops when nothing is happening and wakes on the next message. You are not paying for a machine to sit there overnight — but do expect the first message after a quiet period to be slower than the second.
Not because it is special, but because a bot token from BotFather is the least painful channel setup that exists. All 9 also speak Slack, Discord, email, Matrix, WhatsApp and SMS.
Hosting is fixed and shown in the table. Model tokens are forwarded at provider cost with no markup, and that half is entirely up to how you use it. Read the cost breakdown before pointing one at a long task — a coding agent stuck in a retry loop is the most expensive failure mode on this platform.
Start from the credential you already have. This decides it more often than any feature comparison. An Anthropic key, a ChatGPT plan, a Cursor seat — the runtime that uses what you already pay for is usually the right one, and the Bring column is the fastest way to filter.
Then ask what the output should look like. If you want to review changes as commits, Aider is built around that and nothing else here is. If you want an answer in a chat window, any of them do that.
Then ask how long the tasks are. Short, frequent questions favour pi. Multi-hour work with a plan favours deepagents. Everything in between is Claude Code or Codex territory.
The comparison that does not help much is benchmark scores. All 9 are competent; what determines whether this works for you is whether the key situation is clean, whether the output lands somewhere you can use it, and whether you notice when it goes wrong.
An agent you message from a phone is one you are not watching. Every agent built here reports to ClawMetry — cost per conversation, what tools it ran, and a way to stop it mid-task. For a coding agent with a shell that is the difference between a useful tool and an unattended process with your credentials in it.
This opens the builder with a starting brief already written. Change any of it before you send — first agent month is free.
Start with this brief →