Skip to main content
Podflare ships an MCP server that exposes sandbox operations as tools Claude can call directly from the composer. Once configured, Claude can run Python, execute shell commands, fork sandboxes, and transfer files — all backed by a Firecracker microVM on your Podflare host — without any code changes to your prompts.

​
Supported clients

Claude Desktop

macOS and Windows desktop app by Anthropic.

Cursor

AI-first code editor with MCP tool support.

Cline

VS Code extension for autonomous coding agents.

Zed

High-performance editor with built-in AI assistant.

​
Configure the MCP server

1

Open your MCP client config file

Locate the config file for your client:
~/Library/Application Support/Claude/claude_desktop_config.json
For Cursor, Cline, and Zed, consult their respective MCP configuration docs — the JSON structure below is the same.
2

Add the Podflare server entry

Add a podflare entry under mcpServers. Set the absolute path to the compiled MCP server entry point and supply your Podflare credentials as environment variables:
{
  "mcpServers": {
    "podflare": {
      "command": "node",
      "args": ["/absolute/path/to/podflare/mcp/dist/index.js"],
      "env": {
        "PODFLARE_HOST": "https://api.podflare.dev",
        "PODFLARE_API_KEY": "pk_your_key_here"
      }
    }
  }
}
Replace /absolute/path/to/podflare/mcp/dist/index.js with the real path on your machine and pk_your_key_here with your API key.
3

Restart your MCP client

Quit and reopen Claude Desktop (or your chosen client) to load the new configuration. Podflare tools appear as hammer icons in the composer once the MCP server connects.

​
Available tools

After connecting, Claude can see and invoke the following tools:
ToolDescription
run_pythonREPL-persistent Python execution
run_bashFresh shell for each invocation
forkN-way copy-on-write sandbox fork
diffFilesystem diff between two sandboxes
merge_intoCommit a fork as the new sandbox state
uploadTransfer a file into the sandbox
downloadRetrieve a file from the sandbox

​
Example prompt

You do not need special syntax to invoke tools. Ask Claude naturally:
“Use run_python to load /data/customers.csv with pandas and tell me the top 5 regions by revenue.”
Claude selects run_python, the MCP server routes the call to Podflare, the sandbox executes the code, and the result flows back into Claude’s response.

​
Session model

Each Claude conversation gets one Podflare sandbox. The sandbox is created on the first tool call and destroyed when your MCP client quits the server process. Reopening a conversation starts a fresh sandbox — any in-memory Python state from the previous session is gone.
In-memory REPL state does not survive between Claude Desktop sessions. If you need data to persist across conversations, write it to a file inside the sandbox and use download to save it locally before closing.

​
Troubleshooting

Check the MCP server log at ~/Library/Logs/Claude/mcp-server-podflare.log (macOS). The most common cause is that the Podflare host is not reachable at the address set in PODFLARE_HOST. Verify the host is running and the URL is correct.
The MCP server applies a 5-minute timeout per tool call. Long-running code inside the sandbox can exceed this limit. Use the fork tool to split work across parallel sandboxes and reduce per-call execution time.