How to run Python from Claude or Cursor (code sandbox MCP)

How Plugsky MCP works in practice

Everything in this guide runs on Plugsky’s live MCP endpoint (https://plugsky.com/mcp) or the npm bridge (npx -y @plugsky/mcp). A client request flows: JSON-RPC over Streamable-HTTP → auth (API key or OAuth 2.1 + Dynamic Client Registration with fine-grained scopes) → tool router → the service you asked for. Model calls run behind a 7-tier provider failover chain; browsing renders with headless Chromium; transcripts come from real captions and Groq Whisper; images generate on FLUX; code runs in a sandbox with no network; memory persists in your account (Postgres-backed); and every tool call is metered per user on your dashboard.

Relevant live tools for this article: plugsky_code_run — sandboxed Python: no network, 256MB, 30s, metered.

How to run Python from Claude or Cursor (code sandbox MCP)

Direct answer

To run Python from Claude or Cursor in 2026, install a code-sandbox MCP: E2B (best cold-start), Daytona (persistent envs), Modal, Cloudflare Sandbox, Vercel Sandbox, or Plugsky's plugsky_code_run (roadmap Q4 2026, GPU-optional, sovereign). The MCP spawns an isolated micro-VM, runs your code, and returns stdout + generated files. This is the killer capability behind ChatGPT Advanced Data Analysis — now bring it to any MCP-aware client.

Key facts

Best MCPs today	E2B, Daytona, Modal, Cloudflare Sandbox, Vercel Sandbox
Best cold-start	E2B (<1s)
Best persistent envs	Daytona (dev environments live for days)
GPU-capable	Modal, Cloudflare, Plugsky (Q4 2026)
Auth	Bearer (all), OAuth (Daytona, Cloudflare)
Free tier	Modest on all; self-host E2B open-source for unlimited
Package installs supported	✅ all
Persistent files	✅ (with session/persist flags)
Best for data analysis	E2B + matplotlib/pandas preinstalled

TL;DR

• Code-sandbox MCPs give agents a real Python/Node runtime without leaving the client.
• E2B wins on cold-start speed; Daytona wins on persistent state; Cloudflare/Modal/Plugsky win on GPU.
• Enterprise pattern: use Plugsky Sovereign to run code on your own GPUs — no code leaves your VPC.
• Best pairing: sandbox MCP + web MCP + Plugsky models = a full data-analysis agent.

How it works

Option 1: E2B

{ "mcpServers": { "e2b": { "url": "https://mcp.e2b.dev", "headers": {"Authorization":"Bearer e2b_..."} } } }

Fast cold-start. Great for one-shot analysis.

Option 2: Plugsky plugsky_code_run (Q4 2026)

When shipped, plugsky_code_run gives you sandboxed Python/Node inside the same MCP that already exposes 36 models and your RAG. GPU-optional, sovereign-optional.

Option 3: Daytona for persistent envs

Best when the agent needs a long-lived environment (multi-day project, warm-cached deps).

Winning patterns

• CSV analysis — upload via Plugsky RAG, then plugsky_code_run pandas over it.
• Chart generation — matplotlib output as PNG returned as MCP resource.
• API testing — write curl-equivalents in Python, run instantly.
• Data pipeline prototyping — hit a DB, transform, write CSV back to a collection.
• ML experiments — with GPU-enabled sandbox (Plugsky, Modal), fine-tune small models mid-chat.

Security considerations

• Sandbox isolation: micro-VMs are the safest — kernel isolation, no shared filesystem, ephemeral by default.
• Egress control: enterprise deployments should pin which domains sandboxes can reach.
• Time limits: cap execution time (all sandboxes support this).
• Data leaving your network: use Plugsky Sovereign or self-hosted E2B to keep code+data on-prem.

Comparison — code sandbox MCPs

Cold-start	<1s	~10s	~3s	~1s	~2s	~1s
Persistent envs	⚠️	✅	⚠️	❌	❌	✅
GPU	❌	⚠️	✅	❌	❌	✅
Sovereign / on-prem	Self-host	✅	❌	❌	❌	✅
Free tier	Modest	Modest	Modest	Modest	Modest	Yes (roadmap)
OAuth 2.1	⚠️	✅	⚠️	✅	⚠️	✅

FAQ

Q: Is running Python from Claude safe?

A: The code runs in an isolated micro-VM — not on your machine. Safe as long as you trust the sandbox provider and set egress limits.

Q: Does Plugsky already have plugsky_code_run?

A: On the roadmap for Q4 2026. Today, pair Plugsky's models with an E2B or Daytona MCP.

Q: Can the sandbox access my local files?

A: Not by default — sandboxes are isolated. To share files, upload them via Plugsky RAG (plugsky_documents_ingest) and let the sandbox fetch from the collection.

Q: Can I install pip packages?

A: Yes. Each sandbox supports pip install. First call in a fresh VM is slower; subsequent calls in a persistent env are instant.

Q: What about Node.js?

A: All the top sandbox MCPs support Node. E2B, Modal, and Plugsky all offer Node runtimes.

Q: GPU access?

A: Modal and (upcoming) Plugsky offer GPU sandboxes. Use for ML experiments, embeddings, or vision workloads.

Trust & sources

• Author: Mustafa Hasan.
• Last updated: {DATE}.
• References: E2B docs, Daytona docs, Modal docs.
• Related: [Best MCP servers 2026](/blog/best-mcp-servers-2026) · [Plugsky MCP roadmap](/plugsky-mcp-playbook).