Perplexity SPACE
by Perplexity
Open-source Firecracker microVM sandbox for AI agents, the Apache 2.0 runtime behind Perplexity Computer
Perplexity SPACE (Sandboxed Platform for Agentic Code Execution) is an open-source sandbox platform that runs AI agent code in isolated Firecracker microVMs. It is for platform and security teams that need to self-host agent execution with pause, resume, snapshots and credentials kept outside the sandbox, instead of renting a commercial sandbox service.
SPACE is the sandbox layer Perplexity built for Computer, its long-running agent product, which launched in February 2026 on a commercial sandbox provider and moved onto SPACE in July. Perplexity open-sourced it on October 15, 2026 under the Apache 2.0 licence at github.com/perplexityai/SPACE. Each sandbox is a Firecracker microVM that boots its own Linux kernel with a minimal device model, so agent code is treated as untrusted and does not share a kernel with other tenants the way container sandboxes do. A stateless control plane exposes one HTTP API to create sandboxes from templates, run commands, stream output over server-sent events and manage the lifecycle; lifecycle calls are asynchronous and return an operation ID to poll. Node-level services inject scoped credentials from outside the VM only when needed and enforce egress policy, so an agent never holds long-lived secrets. Storage uses copy-on-write Btrfs, which makes snapshots, pause and resume, and session branching cheap; reports put snapshot retention at up to a week and median creation latency at about 60 ms. The services are written in Rust, and the repo ships a Kind-based local Kubernetes setup, a reference AWS EKS deployment, warm pools, and an E2B compatibility wrapper for code written against the E2B Python SDK. Perplexity says SPACE handles millions of sandbox creations a day, created sandboxes three to five times faster than the provider it replaced, and showed no VM-to-host escape in 108 red-team attempts by frontier models given root, although testers did find network-policy bypasses that were later fixed. Akamai and Crusoe are bringing it to their clouds, and CoreWeave has published setup guidance.
A platform or AI infrastructure team running agents that execute untrusted code at scale and want to own the sandbox layer instead of paying per-sandbox fees to a hosted provider.
MicroVM isolation with snapshots, branching and out-of-sandbox credentials under a permissive licence, already proven on Perplexity's production traffic.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Free
- Target Market
- CTOs, Platform Engineering Teams, AI Infrastructure Engineers, Security Engineers, Enterprise Developers
- Deployment
- Open-source, Self-hosted
Key Features
- ✓Firecracker microVM isolation
Every sandbox boots its own Linux kernel with a minimal device model, giving a stronger boundary for untrusted agent code than containers that share the host kernel.
- ✓Snapshots, pause, resume and branching
Copy-on-write Btrfs storage lets long tasks pause and resume, recover from interruptions and branch into multiple sandboxes without copying whole environments.
- ✓Out-of-sandbox credentials and egress control
Node-level proxies inject scoped tokens only when a request needs them and enforce destination policies, so agents never see API keys or long-lived secrets.
- ✓HTTP control-plane API
One API creates sandboxes from templates, runs commands, streams output over server-sent events and manages lifecycles, with a generated OpenAPI reference.
- ✓E2B compatibility wrapper
Lets code written for the E2B Python SDK run against SPACE, lowering the cost of trying a self-hosted backend, within the documented limits of the wrapper.
- ✓Warm pools and reference deployments
Pre-booted sandboxes on Kubernetes cut start latency, and the repo ships a Kind local setup and a reference AWS EKS deployment to start from.
Capabilities
Use Cases
- •Code execution backend for an internal agent platform
Give every agent task its own microVM to run generated code, collect output and destroy the environment when the task ends, so one bad run cannot reach other tenants.
- •Long-running research and data-analysis agents
Agents that work for hours can be paused, snapshotted and resumed after failures instead of restarting from scratch, as Perplexity Computer does in production.
- •Replacing a hosted sandbox vendor
Teams on E2B can point existing SDK code at a self-hosted SPACE fleet through the compatibility wrapper and compare cost and latency on their own traffic.
- •Agents that call authenticated services
An agent can sign in to an internal or third-party service through node-level credential injection without the token ever being stored inside the sandbox it controls.
- •Managed sandbox offering on a cloud platform
Infrastructure providers can package SPACE as a customer-facing agent runtime; Akamai has launched it on its cloud and Crusoe is adding it to Managed AI Services.
Ideal For
Best For
- ✓Self-hosting code execution for coding, data-analysis or research agents that run untrusted code for minutes or hours
- ✓Teams currently on E2B that want to test a self-hosted backend through the E2B compatibility wrapper
- ✓Long-running agent sessions that need pause, resume, rolling snapshots and branching into parallel sandboxes
- ✓Security-sensitive deployments where agents must reach services without ever holding credentials, under per-sandbox egress policy
- ✓Cloud and GPU providers adding a managed agent-sandbox product, as Akamai and Crusoe are doing
Not Ideal For
- ✗Small teams without infrastructure staff: self-hosting means running a Rust service fleet and Firecracker hosts that need KVM (nested virtualisation or bare metal), plus kernel patching, on-call and ongoing red-teaming
- ✗Teams that want a fully managed service with an SLA today; a hosted sandbox provider trades control for convenience
- ✗Buyers who need stable APIs across every field, since fields marked incubating sit outside the supported contract and the wider Agentic Runtime Platform components are not yet released
Integrations
Deployment
Market Analysis
Pros
- ✓Permissive Apache 2.0 licence with about 1.4k GitHub stars and 113 forks shortly after release
- ✓MicroVM isolation plus snapshotting and branching, with reported median creation of about 60 ms
- ✓E2B compatibility wrapper eases migration from the most common hosted sandbox
- ✓Proven at Perplexity's scale and backed by infrastructure partners Akamai, Crusoe and CoreWeave
Cons
- ✗Heavy to self-host: Rust services, KVM-capable Firecracker hosts, kernel patching and on-call
- ✗Performance and security figures are Perplexity's own and have not been independently reproduced
- ✗Red-teaming found network-restriction bypasses via DNS spoofing and shared-IP routing (since fixed), and 108 failed escape attempts is not proof of no escape
- ✗Partial release: incubating API fields are unsupported and the broader Agentic Runtime Platform is still to come
Pricing
Open source (self-hosted)
$0
- ✓Apache 2.0 licence
- ✓Full control plane, node services and sandbox daemon
- ✓Kind and AWS EKS reference deployments
- ✓E2B compatibility wrapper
SPACE is free under Apache 2.0, so the real cost is the infrastructure and staff to run it: Firecracker hosts with KVM support (bare metal or nested virtualisation), a Kubernetes cluster, and engineers for patching and on-call. Perplexity sells no hosted SPACE plan; managed versions are coming through partners such as Akamai and Crusoe, at their own prices.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 5 on domains other than github.com.
- 1.perplexity.ai — open sourcing space
- 2.github.com — SPACEvendor
- 3.cryptobriefing.com — perplexity open sources space agent sandbox
- 4.techstrong.ai — perplexity unveils space a secure sandbox platform for ai ag
- 5.explainx.ai — perplexity open sources space firecracker agent sandbox apac
- 6.gigazine.net — 20261016 perplexity space
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Gimlet Cloud
Multi-silicon inference cloud that splits AI agent workloads across GPUs, CPUs and accelerators
DigitalOcean Managed Agents
Managed agent runtime with microVM sandboxes, 16,000+ governed tools and serverless inference, billed on active CPU
Modular
MAX inference framework and Mojo language for serving AI models on NVIDIA, AMD and other chips
ZML/LLMD
Free, Python-free LLM inference server that runs open models on NVIDIA, AMD, Google TPU, Intel and Apple chips from one binary