mcp-tenki-sandbox

Give Claude or Cursor a Tenki sandbox (MCP)

Point any MCP client — Claude Desktop, Cursor, Claude Code — at tenki-mcp and your agent gets a disposable Tenki microVM it can drive natively: create sandboxes, run code, read/write files, run git, expose preview URLs, manage snapshots/volumes/templates. 85 tools, full parity with the Tenki API.

No SDK code to write — it's a config entry.

Wire it into your MCP client

Add this to your client's MCP config (Claude Desktop: claude_desktop_config.json; Cursor: .cursor/mcp.json):

{
  "mcpServers": {
    "tenki": {
      "command": "npx",
      "args": ["-y", "@tenkicloud/mcp"],
      "env": { "TENKI_API_KEY": "<your Tenki auth token>" }
    }
  }
}

Your token is the auth_token in ~/.config/tenki/config.yaml after tenki login.

Try it

Once connected, ask your agent:

"Run this Python in a fresh Tenki sandbox and tell me what it prints: print(sum(range(100)))"

It calls tenki_run_code, which boots a microVM, runs the snippet, returns the output, and tears the sandbox down — per-second billed, gone when it's done.

A few of the 85 tools

tenki_run_code	one-shot: boot → run shell/python/js → dispose
tenki_create_sandbox / tenki_exec	a persistent sandbox + run commands in it
tenki_read_file / tenki_write_file	filesystem I/O
tenki_git	clone / commit / push inside the sandbox
tenki_create_preview_url	a public URL for a server the agent starts
tenki_create_snapshot / tenki_create_volume	checkpoint state / attach a persistent disk

Full list: the tenki-mcp README.

Verify

verify.mjs proves the whole path the way an MCP client does: it spawns the server, lists its tools, and drives a real sandbox lifecycle (create → get → terminate) through MCP tool calls against live Tenki.

npm install
node verify.mjs        # needs TENKI_API_KEY (or `tenki login`)
# → ✓ connected — 85 tools advertised
# → ✓ tenki_create_sandbox → …  ✓ tenki_get_sandbox → …  ✓ tenki_terminate_sandbox

It also attempts tenki_run_code as a bonus. That one additionally needs Tenki's per-session data-plane endpoint reachable from your network (it stages the code file there), so it's best-effort — reported as skipped, not failed, if the data plane isn't reachable from where you're running.