CODE-EXECUTION MCP

A code-execution MCP server for your agent.

PillDrops ships a hosted, OAuth-secured MCP server with 70+ tools that spin up microVM sandboxes and run real code. Add one line — no repo to clone, no Docker to babysit — and your agent can create a sandbox, run commands, and read metrics.

Zero install · OAuth on first call · works with Claude, Cursor, any MCP client

add the server — then let the agent run code
$ claude mcp add --transport http pilldrops https://pilldrops.com/mcp
Added HTTP MCP server 'pilldrops' — auth on first call

# agent calls a tool → OAuth (PKCE) in the browser once → done
> create_sandbox(template="python-3.11")
{ id: "sb_9f21", status: "ready" }   # 180ms
> run_command(id="sb_9f21", cmd="python -c 'print(2+2)'")
{ stdout: "4\n", exit_code: 0 }
✓ code executed in a real microVM · billed per second
14
MCP tools
zero
install
OAuth
PKCE on first call
any MCP
client works
$5
free credits
[01] — The toolset / everything an agent needs to run code

70+ tools, one server.

Create a sandbox, move files in and out, run commands, start a server, read logs and metrics, then pause or destroy it. The full lifecycle — no glue code, no DIY Docker wrapper to maintain.

create_sandbox run_command write_file read_file get_logs get_metrics start_server pause resume destroy_sandbox list_sandboxes list_templates get_usage
⬢

Run code

create_sandbox boots an isolated sandbox from a template; run_command executes a shell command and returns stdout, stderr, and exit_code.

⇅

Move files

write_file and read_file put data in and pull results out — datasets, scripts, artifacts — without leaving the tool call.

◷

Serve & observe

start_server exposes a public preview URL; get_logs and get_metrics stream output and CPU/memory back to the agent.

◧

Pause, resume, clean up

pause freezes state (storage-only billing), resume picks it back up, and destroy_sandbox tears it down.

▤

List & discover

list_sandboxes shows what's running and list_templates shows what you can boot — python-3.11, node-20, code-interpreter, and more.

◈

Meter spend

get_usage reports vCPU-seconds, GB-seconds, and cost so the agent — and you — can attribute spend per run. Metering guide →

[02] — Three ways to connect / hosted, local, or CI

Hosted OAuth, local, or CI.

The hosted transport is the zero-setup path: one line, OAuth on first call, nothing to install. Run it locally against your own binary, or wire it into CI with a bearer token.

OPTION 01

Hosted (OAuth, zero install)

Point any MCP client at the hosted endpoint. The first tool call opens a browser for OAuth (PKCE); after that it just works.

$ claude mcp add --transport http pilldrops https://pilldrops.com/mcp
no installOAuth PKCE
OPTION 02

Local (your own binary)

Run the MCP server on your machine over stdio via the PillDrops CLI — handy for air-gapped or self-hosted setups.

$ claude mcp add pilldrops -- pilldrops mcp
self-hoststdio
OPTION 03

CI (bearer token)

In non-interactive environments, skip the browser flow and pass an API key as a header so pipelines can run code in sandboxes.

$ claude mcp add --transport http pilldrops https://pilldrops.com/mcp \
    --header "Authorization: Bearer <api-key>"
headlessCI
[03] — Any MCP client / Claude, Cursor, or your own

Works with any MCP client.

The server speaks the Model Context Protocol, so the same 70+ tools show up in Claude, Cursor, or any client that speaks MCP over HTTP or stdio. Machine-readable capabilities live at /llms.txt.

◆

Claude

Add over HTTP with claude mcp add --transport http, authorize once, and the tools appear inline for the model to call.

▸

Cursor

Register the same hosted endpoint in Cursor's MCP config — no separate build, same OAuth flow, same toolset.

⧉

Any client

Standard MCP over HTTP or stdio means custom agents and other clients connect the same way. MCP reference → · /llms.txt →

ADD IT IN ONE LINE

Give your agent a sandbox.

One command adds a hosted, OAuth-secured MCP server with 70+ tools for running real code in isolated microVMs. No repo to clone, no Docker to maintain. Try it on $5 of free credits — no card.