If you typed "podflare" into Google and ended up on Cloudflare’s homepage, you are not the first. The names are one letter apart, Cloudflare is twenty years older, and Google’s knowledge graph defaults to the entity it already knows. This post is the direct answer.
Podflare is a hardware-isolated cloud sandbox for AI agents. We are not part of Cloudflare, affiliated with Cloudflare, or a Cloudflare product. We are a separate, independent company building a separate product. The similarity of the name is a coincidence (and, in retrospect, a product of wanting a short four-syllable domain that was still available).
What Podflare actually does
An AI agent that writes and runs code needs somewhere safe to execute it. Your own servers are the wrong place — one prompt-injected rm -rf away from disaster. Serverless functions (Lambda, Cloud Run) are too restrictive — no persistent state between calls, no full filesystem access, time limits in the seconds. Docker containers share a kernel with the host, and the last few years of kernel CVEs have shown that’s not a strong enough boundary when the code was written by an LLM responding to untrusted user input.
Podflare is the right place. Every time your agent calls Sandbox(), we hand it a brand-new, dedicated Linux microVM:
- Hardware-isolated. Each VM is a Podflare Pod — a dedicated KVM-backed microVM, the same category of isolation AWS Lambda is built on. The security boundary is the hypervisor, not the kernel. Container-escape CVEs don’t apply.
- Fast. Create + exec + close round-trips at ~190 ms p50 from a residential laptop (p99 under 240 ms), ~43 ms p50 from an agent running in a nearby cloud region.
- Persistent. Each sandbox keeps a live Python REPL across
run_codecalls. When your agent imports pandas on turn 3 and uses it on turn 7,globals()still has the import. - Forkable.
parent.fork(n=5)spawns 5 children from the parent’s exact mid-flight state in ~80 ms. The primitive tree-of-thought and multi-attempt code synthesis patterns actually want. - Multi-region. 5 production regions today — us-west, us-central, us-east, eu, sg — with automatic haversine routing from the edge and failover on origin 5xx.
The one-paragraph architecture
When you call Sandbox(), the request hits a Cloudflare Worker (yes, we use Cloudflare — as a CDN, like everyone — but we are not Cloudflare) that haversine-routes to the nearest region. That region runs hostd, our Rust-based control plane, which pulls a pre-booted Podflare Pod microVM off a warm pool (6 ms server-side), assigns it to your org, and returns the sandbox ID. Your run_code calls stream over a vsock binary protocol inside the guest — no TCP, no TLS inside the VM, which is why our hot-exec is 46 ms p50 vs E2B’s 180 ms. The full benchmark post has the numbers.
What you’d use Podflare for
The common patterns we see:
- Code-execution tool for an AI agent. Claude / GPT / Gemini / Llama emits Python as a
tool_useor function call; your app routes it to Podflare; the stdout comes back; the agent uses it on the next turn. - Tree-of-thought solver. Load an expensive context once (a large DataFrame, a trained model), then
fork(n=5)to try five strategies in parallel. Pick the winner, merge it back, destroy the losers. - Long-running agent sessions. Create a
persistent=Truesandbox, train a model, load a dataset, then freeze it to a Space when idle. Resume hours or days later with the same Python process still holding the same variables. - Drop-in replacement for OpenAI’s code interpreter — except you control the infrastructure, the data doesn’t go to OpenAI, and your bill isn’t tied to a model-tier quota.
Pricing
Free tier includes a $200 starter credit, 10 concurrent sandboxes, 1 GB RAM per sandbox, 5 regions, and the full SDK. Pro is $29/month (50 concurrent, 4 GB), Scale is $200/month (500 concurrent, 16 GB), Enterprise is custom with SOC-2-roadmap access, dedicated hostd pools, and data residency. Full breakdown on the pricing page.
How to try it in 60 seconds
pip install podflare # or: npm install podflare
export PODFLARE_API_KEY=pf_live_... # mint one at dashboard.podflare.ai/keys
python -c "
from podflare import Sandbox
with Sandbox() as sb:
out = sb.run_code('print(2 + 2)')
print(out.stdout)" # 4That’s the whole onboarding. Real compute, real network, real hardware isolation, real sub-200 ms cold start. Create a free account and run the above.
Related reading
- Podflare vs Cloudflare: different companies, different products — the full disambiguation, for the long version.
- Cloud sandbox benchmark: E2B vs Daytona vs Podflare — the numbers we use to back up the "fastest" claim.
- What is a cloud sandbox for AI agents? — the category overview, if you’re evaluating all players.