Demo
Freeno card
Self-serve. Kick the tires on real microVMs.
- 2 concurrent sandboxes, 1 hour each
- Live-fork, exec, suspend / resume
- Community support
Fork your RL environments.
Don't restore them.
Collimate warms a microVM once, then forks the running VM — process memory and all — into thousands of rollouts in milliseconds. No cold boot between episodes, no snapshot restore tax. The same engine gives coding agents a real, isolated shell.
No card. 2 concurrent sandboxes on the free Demo tier.
$ curl -sX POST api.collimate.ai/v1/sandboxes \ -H "Authorization: Bearer $COLLIMATE_API_KEY" -d '{"template":"python"}' { "id": "sbx_a1f09c", "state": "running" } $ curl -sX POST api.collimate.ai/v1/sandboxes/sbx_a1f09c/exec \ -d '{"command":"python setup_env.py"}' # run the expensive prefix once { "exit_code": 0, "stdout": "env ready at step 0" } $ curl -sX POST api.collimate.ai/v1/sandboxes/sbx_a1f09c/fork \ -d '{"count":64}' # fork the live VM into 64 rollouts { "parent": "sbx_a1f09c", "children": [ ... ] } # process memory preserved, no replay $ curl -sX POST api.collimate.ai/v1/sandboxes/sbx_a1f09c/suspend # stops the meter { "id": "sbx_a1f09c", "state": "suspended" }
Prepare once. Fork many.
Everything is tuned for the part of the post-training loop that runs millions of times: spawning, branching, and tearing down isolated environments without the environment becoming your bottleneck.
Fork
Fork a running microVM in single-digit milliseconds. The clone keeps live process memory, so RL rollouts and tree search explore many continuations from one expensive prefix instead of replaying it.
Schedule
Warm-pool and affinity placement keep sandboxes off your trainer's critical path, built for GRPO-scale rollouts and agentic fan-out alike.
Store
Suspend to snapshot storage between turns and resume with memory intact. You pay per second while live and per GB-month while suspended, never per sandbox.
One engine, two workloads.
The same live-fork primitive that fans out rollouts gives a coding agent a contained place to run.
RL post-training
Fan out GRPO-scale rollouts and fork to a mid-episode state instead of replaying it. Hold a VM alive as a session and drive it across many steps, the exact shape an RL environment's step loop needs.
Coding agents
Give an agent a real shell and Python per task, over REST, the Python SDK, or hosted MCP. Every sandbox is a hardware-isolated microVM, so model-generated code and tool calls run contained.
Three tiers, one engine.
Most teams
Pro
Usageno platform fee
Production fan-out, billed per second.
- Up to 100 concurrent sandboxes, 24h sessions
- Per-tenant keys and private templates
- Suspend stops the meter; forking is free
Enterprise
Custominvoice
Dedicated capacity, SSO, or your own cloud.
- Unlimited concurrency and session length
- Dedicated nodes, SSO, BYOC
- Onboarding, SLA and security review
vCPU $0.040 / hr · RAM $0.012 / GB-hr · suspended storage $0.10 / GB-mo · spawning and forking free. See full pricing →