zode mcp runs an MCP server on stdio. Your agent starts it as a subprocess and gets nine tools for sandboxes. It acts with your zode sign-in, in your workspace, with your plan’s limits.
Sign in first
The server reads the same credentials as the CLI: the API key saved byzode login, or the API key in ZODE_TOKEN when set. With neither, it exits with not signed in: run zode login first, then start your agent.
Add it to your agent
- Claude Code
- Codex
- Cursor
ZODE_TOKEN in its environment. In the Cursor config, add "env": { "ZODE_TOKEN": "zode_tok_..." } next to args.
The server also sends the agent short instructions: work happens in /workspace, refer to sandboxes by name, fork to try two approaches, delete when done.
Agent skill
A skill teaches Claude Code, Codex, Cursor and other agents when and how to use Zode: making sandboxes, forking for parallel work, snapshots, images and secrets. It is built from these docs, so it stays current.Tools
Everysandbox parameter takes a sandbox name or sbx_ id. Tools that return a sandbox return name, id, status, size and region.
sandbox_create
sandbox_exec
It returns
exit_code, stdout, stderr, duration_ms, and timed_out or truncated when they apply. Each stream is capped at 256 KB; past that, output is dropped and truncated is true. A timed-out command returns exit_code -1. For a long-running server, start it in the background (nohup ... &).
Paths
sandbox_read_file and sandbox_write_file take an absolute path, or a path relative to /workspace. sandbox_read_file returns the file in stdout, with the same fields as sandbox_exec.
Limits
sandbox_createhas no environment or spend-cap input. For those, usezode newwith--environmentor--max-costand give the agent the name.- The tools need a running sandbox for exec and files. A stopped one gives
demo is stopped; start it first. - Plan limits apply as in the CLI: on Free, at most 2 sandboxes run at once, and a workspace with no card and no credit cannot make sandboxes.