Get started

Bugnet gives your AI agent a throwaway cloud Linux sandbox for work you don't want on your computer: a repo you have not checked, an AI model, a tool your machine lacks. Your agent pays per run in USDC on Solana. There is no account and no API key.

sequenceDiagram
    participant You
    participant Agent
    participant Bugnet
    participant Sandbox
    You->>Agent: Try this repo
    Agent->>Bugnet: bugnet_assess
    Bugnet-->>Agent: findings, price
    Agent->>You: findings, price
    You->>Agent: approve
    Agent->>Bugnet: bugnet_start
    Bugnet-->>Agent: pay link, claim
    You->>Bugnet: pay in USDC
    Agent->>Bugnet: bugnet_redeem
    Bugnet->>Sandbox: start
    Agent->>Bugnet: bugnet_exec, files
    Bugnet->>Sandbox: run, read, write
    Agent->>Bugnet: bugnet_stop
    Bugnet->>Sandbox: delete

1. Install Bugnet#

Run this in a terminal to add Bugnet to Claude Code:

claude mcp add --scope user --transport http bugnet https://mcp.bugnet.sh/mcp

--scope user puts Bugnet in every project, for you only. It needs no sign-in, header or API key. By installing Bugnet you accept the Terms.

Check it with claude mcp list. The list shows bugnet: https://mcp.bugnet.sh/mcp (HTTP) - ✔ Connected.

This entry pays by link. To pay from the pay wallet on this computer, add the bridge instead, which needs Node 20.3 or newer. Add one, not both, or every tool shows up twice:

claude mcp add --scope user --transport stdio bugnet -- npx -y @bugnet/mcp@0.1.0

The Claude Code plugin adds a skill and hints on top of the server. Every other client is on Install.

2. Get ready to pay#

Bugnet takes USDC on Solana mainnet only, and pays the SOL fees itself. Set up one way to pay:

To set up the pay wallet, about 5 to 10 minutes once:

1. Install pay with brew install pay, or with npm i -g @solana/pay.
2. Run pay setup.
3. Run pay topup.
4. Start Claude Code as plain claude, never pay claude, which loads only pay's own server and hides Bugnet.

brew install pay
pay setup
pay topup

A machine with no biometric or polkit prompt, such as over SSH or headless, cannot pay this way.

Refunds go back automatically to the USDC token account that paid, so keep it open. Before you pay compares the two ways.

3. Ask your agent#

Ask in your own words. For example:

Try https://github.com/someone/some-tool (opens in a new tab) for me, I don't trust it.

Or "get this text to speech model running". Your agent learns from Bugnet's server instructions when a sandbox fits: AI models, code you have not checked, tools your machine lacks. It offers a sandbox once, never for work on your own project, and does not ask again for a task once you say no.

4. The free check#

Your agent calls bugnet_assess, which is free. It checks the repo, model or script against fixed rules:

• the repo's age and stars;
• install scripts that download or run binaries;
• hidden code;
• model files that can run code when loaded;
• whether your machine has the memory or GPU the work needs.

bugnet_assess returns short findings, never the files it read, and a price for the sandbox sizes that fit. bugnet_quote gives the prices without the check.

The agent tells you in a sentence or two what it found and the price, for example that the repo is 3 days old and its installer downloads a binary. Then it waits for your answer. Nothing is paid yet, and the quote is valid for 10 minutes.

5. Approve the price#

1. Say yes to the agent.
2. Read the price in Claude Code's approval prompt for bugnet_start.
3. Approve the call.

The agent calls bugnet_start with confirmPriceUsd, the price it told you, so the price is in the prompt. Bugnet refuses the call, before any order exists, if that price is not the exact price of the size the agent picked. Approving means "yes, pay this exact amount for this sandbox".

Claude Code asks before every bugnet_start, even in modes that skip other prompts, and refuses the call when there is no person to ask.

Bugnet answers with a payment step: an order, its pay link on pay.bugnet.sh, and a claim that lets your agent use the order once it is paid.

6. Pay#

The amount is the price you approved. An unpaid order expires after 15 minutes, and nothing is charged.

By link#

1. Open the pay link the agent shows you, https://pay.bugnet.sh/o/<order>.
2. Pay with your browser wallet, or scan the page's QR code to open it in Phantom on your phone.
3. Tell your agent you paid.

From the pay wallet#

1. Approve the payment with Touch ID when pay asks.

The bridge pays only links on pay.bugnet.sh, only on Solana mainnet, and never more than the price you approved. If pay is not set up, the bridge opens the pay link in your client when the client can open links, and waits up to 10 minutes for you to pay there.

7. The sandbox starts#

After a link payment, your agent calls bugnet_redeem with the order and its claim. With the pay wallet, the bridge finishes the order itself. bugnet_redeem is free and asks for no approval, because your wallet signature was the payment approval.

Bugnet starts a fresh sandbox, and your agent gets two things:

• The handle, which starts with sbx_. It is the only key to your sandbox, so keep it private.
• The session page link, with the line "After this sandbox ends, its summary and logs stay at <url> for 48 hours", and an instruction to show you the link.

Paid time starts when the sandbox is ready, not when you pay. bugnet_status shows the state and the time left.

8. Run commands#

1. Ask for the work, such as "install it and run the tests".

The agent runs each shell command with bugnet_exec, which returns the exit code and the last 8 KiB of the output and of the errors. The agent reads the results and tells you what happened.

• Every tool call answers within 45 seconds. A longer command keeps running, and the agent follows it with bugnet_job.
• bugnet_cancel_job stops a running command.
• The sandbox runs Linux. Windows and macOS programs can be inspected there but not run.
• Bugnet strips terminal escape codes, hidden control characters and text direction tricks from the output, and labels it as data, not instructions.

9. Move files in and out#

bugnet_list_files lists a folder. bugnet_transfer makes a one-use curl link, valid for 10 minutes, that checks the file's SHA-256 both ways. Claude Code asks you before each transfer, since it moves a file between your computer and the sandbox. Downloads go to ./bugnet-downloads/ unless you name another file, and never overwrite a file you have.

Download what you need before the deadline.

10. Watch the run#

1. Open the session page link the agent showed you.

While the sandbox runs, the page shows its output live. After it ends, the page shows:

• a map of the calls Bugnet saw, with Bugnet's own payments, their amounts and transaction signatures linked to an explorer;
• each command's output and errors, with exit codes;
• the start time, the end time and why it ended.

The page is read-only, and anyone with the full link can read it. Watching a run has the rest.

11. Add time#

1. Ask the agent for more time.
2. Approve the bugnet_extend prompt.
3. Pay the same way as in step 6.

bugnet_status shows the time left and what more time costs. bugnet_extend is paid and asks you every time, like bugnet_start. After a link payment the agent finishes it with bugnet_redeem. Extend before the deadline: an ended sandbox cannot get more time.

12. Stop#

1. Download what you need.
2. Tell the agent you are done.

The agent calls bugnet_stop, which deletes the sandbox and everything in it at once. Its answer carries the session link again, with the time the page is deleted: "Summary and logs: <url> (until <time>)".

If nothing stops it, the sandbox and everything in it are deleted at the deadline. Unused time is not refunded when you stop early.

13. If something goes wrong#

Refunds are automatic and go back to the USDC token account that paid. The full list is in the Terms.

At the end you have the results in your chat, the files you downloaded, and the session page for 48 hours. The sandbox and everything in it are gone.