
Hardware-isolated cloud sandboxes for AI agents, in private alpha
Podflare suits teams that run untrusted LLM-generated code and want a persistent Python REPL and forking. It is in private alpha, holds no compliance certification and relies on vendor-run benchmarks from April 2026. Pick E2B or Daytona if you need an open-source core, a larger free sandbox or an audited vendor.
Fork primitive: tree-search AI agents can branch any running sandbox in under 100 ms without restarting, a capability not found in E2B or Daytona.
Speed benchmarks and a unique fork primitive create technical differentiation, but the sandbox market is contested and well-funded competitors like E2B and Cloudflare can replicate the approach with sufficient investment.
Main trade-off
No GPU plans, no compliance certifications, no published uptime SLA below Enterprise, and limited public information about team or funding creates uncertainty for production commitments.
| Plan | CPU | RAM | Storage | Price |
|---|---|---|---|---|
| FreeCloud · Prototyping with up to 10 concurrent sandboxes, a 30-minute max lifetime and a 5-minute idle timeout. Includes a $200 new-account credit. · US-West · Europe | 2 vCPU | 1 GB | 4 GB SSD | $0/mo |
| ProCloud · Production workloads with up to 50 concurrent sandboxes, an 8-hour max lifetime, a 30-minute idle timeout and a 24-hour priority support SLA. · US-West · Europe | 4 vCPU | 4 GB | 16 GB SSD | $50/mo |
| ScaleCloud · High-throughput pipelines with up to 500 concurrent sandboxes, a 24-hour max lifetime, a 2-hour idle timeout and a 4-hour support response. · US-West · Europe | 8 vCPU | 16 GB | 64 GB SSD | $200/mo |
Free: $0 with a $200 new-account credit; per sandbox 1 GB RAM, 2 vCPU, 4 GB disk, 30-minute max lifetime, 5-minute idle timeout, 10 concurrent sandboxes. Pro: $50/month; 4 GB RAM, 4 vCPU, 16 GB disk, 8-hour lifetime, 30-minute idle timeout, 50 concurrent, 24-hour support SLA. Scale: $200/month; 16 GB RAM, 8 vCPU, 64 GB disk, 24-hour lifetime, 2-hour idle timeout, 500 concurrent, 4-hour support response. Enterprise: custom, unlimited session length, dedicated host pool. Overage on Pro and Scale: $0.05 per vCPU-hour, $0.016 per GiB RAM-hour, $0.09 per GB egress, metered per second and billed monthly. In private alpha, paid upgrades are activated by email within 15 to 30 minutes with no card hold. One vendor comparison table lists the starter credit as none; the pricing page, read 2026-10-26, says $200. Different limits apply at Daytona, a rival listed in our directory.
$0.09 per GB egress, metered per second and billed monthly. No inclusive bandwidth on any plan tier.
Podflare runs AI agent code inside KVM microVMs, each with its own kernel, and remains in private alpha. In its own April 2026 tests the vendor reports a 153 ms median create-and-run cycle, with a free tier and paid plans available by email upgrade.
KVM Hardware Isolation
Each sandbox runs inside a KVM-backed microVM with a dedicated kernel, preventing cross-tenant code leaks that container-based isolation cannot block.
46 ms First-Exec Latency
Inside a freshly created sandbox the first code execution runs in about 46 ms on Podflare's April 2026 benchmark, well ahead of E2B and Daytona in the vendor's own tests.
Fork Primitive
A running sandbox can be snapshotted and branched into several copies; the docs report about 80 ms at the median for a five-way fork, so tree-search agents can explore parallel paths.
Persistent Python REPL
Variables, imports and loaded data survive across run_code calls, and data-science templates preload pandas, numpy, scipy and matplotlib.
Interactive PTY Sessions
Added in April 2026, full pseudo-terminal support lets agents run interactive tools like htop, npm init or any REPL that needs terminal control sequences.
Egress Control
Outbound internet is on by default, and Sandbox(egress=False) detaches the network so untrusted code cannot call out; forked children inherit the setting.
Persistent Spaces
Opt-in persistent sandboxes freeze full VM memory to disk for cross-session resume, stored only in the region that created them.
| Service | Regions | Pricing | Compliance | Checked |
|---|---|---|---|---|
| US · Europe | Free tier · Monthly · Pay as you go | -- | 26 Oct | |
| -- | -- | -- | -- | |
| US · Europe | Pay as you go | SOC 2 · HIPAA | 4 Oct | |
| -- | Free tier · Subscription · Monthly | -- | 22 Nov |
Podflare gives AI agents hardware-isolated Linux microVMs for running code, with a persistent Python REPL, a fork primitive and an MCP server for coding assistants. It is still in private alpha as of 26 October 2026: paid upgrades are activated by email, and no compliance framework is certified. SOC 2 Type II was targeted for Q3 2026 and is not yet held, so regulated workloads should wait. EU-only routing is available today for data residency.
Free costs $0 and adds a $200 new-account credit. Pro is $50 per month with 4 GB RAM and 50 concurrent sandboxes; Scale is $200 per month with 16 GB RAM and 500 concurrent sandboxes; Enterprise is custom. Overage beyond plan limits is $0.05 per vCPU-hour, $0.016 per GiB RAM-hour and $0.09 per GB of egress, metered per second and billed monthly.
Yes. A new account gets the free tier plus the starter credit, which the pricing page says applies to both subscription and usage fees. Free sandboxes are capped at 1 GB RAM, 2 vCPU and a 5-minute idle timeout, so long jobs need an upgrade. Upgrades are enabled by email rather than a checkout button.
E2B is the closest rival: it also runs microVMs, has an open-source core and offers a larger free sandbox. Daytona uses Docker with Sysbox and is the better fit for GPU or Windows needs. Blaxel offers pause and resume, and Modal suits teams that want GPU functions. Podflare's own comparison page lists the first three, so weigh its claims as vendor-supplied.
On Podflare's own April 2026 test of repeated sequential runs, a create-and-run cycle took 153 ms at the median versus 467 ms for E2B, and first execution inside a ready sandbox was faster still. The vendor ran and published both tests. Choose Podflare for latency-sensitive or fork-heavy agents; choose E2B for a bigger free tier and open-source licensing.
Podflare is a cloud sandbox for AI agents, built by Robel Tegegne, who is named as founder on the company blog that starts in April 2026. It is still in private alpha: the pricing page says paid upgrades are activated by email instead of through self-serve checkout. Each sandbox is a KVM-backed microVM (a small virtual machine with its own Linux kernel), which is the main difference from container sandboxes that share a host kernel. Direct rivals include E2B, plus Daytona, which is Docker-based. The company has no connection to Cloudflare despite the similar name.
Three features define the product. Fork snapshots a running sandbox and starts copies from that exact state, which the vendor times at roughly 80 ms for five children (the homepage rounds that to 100 ms). A persistent Python REPL keeps variables, imports and loaded data alive between calls, so an agent does not reload a DataFrame every turn. Interactive PTY sessions, added in April 2026, let agents drive tools such as htop or npm init. Adapters exist for the OpenAI Agents SDK, the Vercel AI SDK, Anthropic tool use, LangChain and Gemini, plus an MCP server, so coding assistants such as Claude Code, Cursor and Cline can run shell commands in a sandbox instead of on a laptop.
It fits teams building code-interpreter agents on Claude or OpenAI models, and teams that treat LLM-generated code as untrusted. Outbound internet is on by default and can be switched off per sandbox. It does not fit GPU inference, managed databases or Kubernetes; for GPU work look at Modal or browse the cloud services directory.
The speed figures come from Podflare's own April 2026 benchmark of 30 sequential create-and-run cycles from a US laptop, which put Podflare ahead of E2B, Blaxel and Daytona at the median and in the tail. First execution inside a fresh sandbox is a separate, much smaller figure. Both are vendor-run results, and the site cites no independent benchmark.
The case against Podflare today is maturity. Its own comparison table shows a smaller free sandbox than E2B Hobby (1 GB RAM against 8 GB), and E2B has an open-source core while Podflare's server is proprietary with an MIT-licensed SDK. The newest blog post is dated 24 April 2026, and regions are described inconsistently: the pricing and security pages name US-West and Europe, while llms.txt claims five. Teams that need an audited vendor should wait, because no compliance audit is complete and the Q3 2026 SOC 2 target has passed. Not sure which sandbox fits your agent? Try Smart Match.
Plans, regions, the uptime commitment and the compliance list on this page are read from Podflare's own pages and documentation. The verdict is Hokai's reading of them, not the provider's claim.