Fork
Live-fork a running sandbox into N copy-on-write children in milliseconds. The clone keeps the parent's live process memory, so there is no reinstall, no reboot, and no replay tax to reach a mid-episode state.
Fork, Schedule, Store
Fast live-fork microVM sandboxes. Prepare a sandbox once, then fork it into many — each clone keeps the parent's live process memory, so there is no cold boot and no replay tax. One engine for agents and for RL post-training.
$ curl -sX POST api.collimate.ai/v1/sandboxes -d '{"template":"python"}'
{ "id": "sbx_9f2a1c7b3e40", "state": "running" }
$ curl -sX POST .../sandboxes/sbx_9f2a1c7b3e40/exec -d '{"command":"pip install numpy"}'
exit 0 # prepare the sandbox once
$ curl -sX POST .../sandboxes/sbx_9f2a1c7b3e40/fork -d '{"count":8}'
forked 8 children, each keeps the parent's live memory
sbx_1a4f… sbx_2b5e… sbx_3c6d… sbx_4d7c… +4
# fork ~2.4 ms p50 — no reinstall, no reboot, no replayWhat Collimate is
Section titled “What Collimate is”Collimate is a managed sandbox cloud built on hardware-isolated Firecracker microVMs. Every sandbox is a real virtual machine with its own kernel — the same isolation boundary that runs AWS Lambda and Fargate — not a shared-kernel container.
What sets it apart is how a sandbox starts and multiplies. Instead of cold-booting a VM per request, Collimate keeps a warm parent resident and forks it copy-on-write on demand, so a fresh isolated sandbox is ready in single-digit milliseconds. Because the fork is a live fork, the child inherits the parent's running process memory (installed toolchain, loaded model, warmed cache). You prepare once and fork many instead of rebuilding state in every clone.
Fork · Schedule · Store
Section titled “Fork · Schedule · Store”Schedule
The platform keeps capacity ready and following your live demand, so a create or fork starts instantly instead of waiting on a cold boot.
Store
Content-addressed, layer-preserving templates hold hundreds of environment variants for close to the cost of one base image, and move only the thin diff when distributing a new variant.
Two audiences, one engine
Section titled “Two audiences, one engine”Agents
Give an agent a real shell and Python per task, over REST, the Python SDK, or the hosted MCP endpoint. Create, exec, fork the running sandbox, suspend, resume, snapshot. Prepare an environment once and branch it to explore many actions.
RL / post-training
Drive the same fleet from the collimate-rl SDK for rollout
workloads: per-fork egress and
deterministic replay so a graded rollout reproduces
bit-for-bit. Fan out a GRPO group into hundreds of isolated VMs without
starving your GPUs on cold environments.
Why live fork
Section titled “Why live fork”Search and rollout are branch-heavy. From one prepared state an agent wants to try many next actions; a best-of-N sampler wants K candidates from the same prompt; a tree search wants to expand many children of one node. The naive way to get each branch its own isolated environment is to boot a fresh VM and replay the whole history to reach the branch point — the replay tax. It grows with episode length and dominates long-horizon work.
Live fork removes it. Reach the branch point once, then fork the running sandbox into as many children as you need. Each child is a full, isolated microVM that resumes from the parent's exact live memory — including processes, open files, and caches — in milliseconds. Read more in Live fork.
Get started
Section titled “Get started” Authentication Bearer col_ keys, tiers, and 401/429 semantics.
Sandboxes quickstart Create, exec, and fork a sandbox over REST or the Python SDK.
MCP quickstart Add the hosted MCP endpoint to Claude Code or Cursor.
RL quickstart Drive rollouts with the collimate-rl SDK.
Live fork The headline: fork a running sandbox, no replay tax.