forkd

What is forkd?

forkd is an open-source sandbox runtime for AI agent fan-out, built on Firecracker and KVM. A parent microVM boots once, loads your runtime such as Python with its dependencies or a model, and is paused to disk as a snapshot. Each child then maps that memory image copy-on-write, so children share resident memory with the parent until they diverge and start far faster than a cold boot. The README reports 100 microVMs forked in about 101 ms and a live branch of a running sandbox with a source pause around 56 ms. It also supports branching mid-run, stacked diff-snapshot chains, per-child network namespaces and cgroup memory limits. You drive it through a CLI, a REST API, a Python SDK or an MCP server, and the Python SDK is described as a drop-in for the E2B Sandbox import. It is licensed under Apache 2.0 and has no paid tier, but you host it yourself on Linux with KVM.

Use cases

• Fan out parallel agent attempts
• Branch an agent mid-task
• Run untrusted code in isolated VMs
• Run code-interpreter sandboxes at scale
• Evaluate agent rollouts in parallel
• Cache installed packages as snapshots

Who it's for

• Agent platform engineers
• Infrastructure engineers
• Rust developers

Why use forkd

• Parallel agent attempts start quickly because children come from a warmed parent instead of cold-booting a kernel.
• An agent can branch mid-thought and try different steering hints while keeping the same filesystem and reasoning state.
• Each child is its own KVM microVM, so isolating untrusted code does not rely on container boundaries.
• Copy-on-write memory keeps the per-sandbox overhead low, reported as 0.12 MiB per child at 100 sandboxes.
• Apache 2.0 licensing with no vendor SDK lock-in, and a Python SDK shaped like E2B's.

forkd features

• Fork many Firecracker microVMs from one warmed parent snapshot
• Branch a live sandbox, with a live mode that shortens the source pause
• KVM hardware isolation for each child
• Copy-on-write memory shared between parent and children
• Stacked diff-snapshot chains for cached package layers
• Per-child network namespace and cgroup v2 memory limit
• REST API, CLI, Python and TypeScript SDKs, and MCP server
• Prometheus metrics and an append-only JSON audit log
• Starter Kubernetes manifest for deployment

forkd integrations

• Firecracker
• LangGraph
• AutoGen
• CrewAI
• Kubernetes
• Prometheus
• Claude Desktop
• Claude Code
• Cursor
• Cline

How to get started with forkd

1. Confirm your Linux host has KVM, then run forkd doctor to check the prerequisites.
2. Boot a parent microVM, import the runtime your agents need, and pause it to a snapshot.
3. Fork the number of children you want from that snapshot through the CLI, REST API or SDK.
4. Branch a running sandbox when an agent needs to explore alternatives from its current state, then collect the results.

forkd limitations

• Needs Linux 5.7 or newer with KVM and a vendored Firecracker fork for live branching
• Self-hosted: you run the daemon, snapshots and nodes yourself
• Managed Kubernetes typically needs metal nodes or nested virtualization
• Benchmark figures come from the project's own tests on its own host

forkd pricing

forkd is open source, so you can use and self-host it for free.

Prices can change and may vary by country or exclude tax. Check the official pricing page.

forkd alternatives

The main forkd alternatives listed on Debuly are E2B, CubeSandbox, Modal, Hyperbrowser, LangChain and Runloop. The table compares pricing and platforms; order is A–Z.

1. 

   CubeSandbox

   Self-hosted MicroVM sandbox service for AI agents, with fast starts, hardware isolation and E2B SDK compatibility.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit CubeSandbox (official site)
2. 

   E2B

   Cloud sandboxes where AI agents run code safely, with Python and JavaScript SDKs and per-second usage billing.

   Developer platformsAutomation and agentsCoding
   FreemiumVisit E2B (official site)
3. 

   Hyperbrowser

   Cloud infrastructure that gives AI agents hosted Chrome sessions, isolated sandboxes, and web fetch, crawl and search APIs.

   Developer platformsAutomation and agentsCoding
   FreemiumVisit Hyperbrowser (official site)
4. 

   LangChain

   Open-source framework for building LLM agents, paired with LangSmith for tracing, evals and agent deployment.

   Developer platformsAutomation and agentsCoding
   FreemiumVisit LangChain (official site)
5. 

   Modal

   Serverless cloud platform for running AI inference, training, batch jobs and agent sandboxes, billed per second of compute.

   Developer platformsAutomation and agentsCoding
   FreemiumVisit Modal (official site)
6. 

   Runloop

   Sandboxed devboxes, blueprints and snapshots for building and running AI software engineering agents.

   Developer platformsCodingAutomation and agents
   FreemiumVisit Runloop (official site)

forkd FAQ

Sources & verification

Facts on this page come from the official sources below, each with the date we read it. Written by Debuly; the vendor did not review this page.

Image credits

• Logo: from avatars.githubusercontent.com · fetched 24 Oct 2026 · Vendor trademark, used only to identify the product; removed on request
• Cover: none collected yet, so a generated tile is shown.

Spot an error? Report an issue with this listing. We also remove a logo or image on the owner's request.