# Collimate

> Fast live-fork microVM sandboxes for AI agents and RL post-training. Every
> sandbox is a hardware-isolated Firecracker microVM. The headline capability is
> LIVE FORK: prepare a running sandbox once, then fork it into N copy-on-write
> children that each keep the parent's live process memory — no cold boot, no
> reinstall, no replay tax. One engine, two audiences (agentic API/MCP, and the
> collimate-rl SDK for rollouts).

Base API URL: https://api.collimate.ai  ·  Hosted MCP: https://mcp.collimate.ai
Auth: `Authorization: Bearer col_<tier>_<id>_<secret>`. Tiers: demo, pro, ent.

## Start here (agents first)

- [Skills for coding agents](https://docs.collimate.ai/agents/): the highest-leverage entry — how to make an agent productive with Collimate.
- [Collimate agent skill (SKILL.md)](https://docs.collimate.ai/skills/collimate/SKILL.md): drop-in skill file teaching an agent the API + MCP + the prepare-once-fork-many workflow.
- [MCP quickstart](https://docs.collimate.ai/start/mcp/): add the hosted MCP endpoint to Claude Code or Cursor.

## Concepts

- [Introduction](https://docs.collimate.ai/): Fork / Schedule / Store, two audiences one engine.
- [Concepts](https://docs.collimate.ai/concepts/): sandboxes, fork-on-demand, live fork, warm pool, templates.
- [Live fork](https://docs.collimate.ai/core/live-fork/): the replay-tax argument and fork-tree search — the headline.

## Getting started

- [Authentication](https://docs.collimate.ai/start/authentication/): col_ bearer keys, tiers, 401/429 semantics.
- [Sandboxes quickstart](https://docs.collimate.ai/start/sandboxes/): create, exec, fork — REST / Python / MCP.
- [Pro quickstart: GRPO rollouts](https://docs.collimate.ai/start/grpo/): bake your own image, fork at rollout width, session lifecycle (client.env).
- [RL quickstart](https://docs.collimate.ai/start/rl/): the collimate-rl SDK (connect, Sandbox, graders, Fleet).

## Core

- [Sandbox lifecycle](https://docs.collimate.ai/core/lifecycle/): running / suspended / deleted and the meter.
- [Running code (exec)](https://docs.collimate.ai/core/exec/): command / code / commands / files / timeout_seconds / seed / egress.
- [Suspend & resume](https://docs.collimate.ai/core/suspend-resume/): pause one sandbox, stop the meter.
- [Files](https://docs.collimate.ai/core/files/): byte-exact upload/download over the exec channel.
- [Templates](https://docs.collimate.ai/core/templates/): the image a sandbox forks from + warm-pool policy.
- [Deterministic replay](https://docs.collimate.ai/core/replay/): pin the guest CRNG seed; seed-ack verdict.
- [Per-fork egress](https://docs.collimate.ai/core/egress/): default-deny allowlist; metadata endpoint always blocked.

## Reference

- [API reference](https://docs.collimate.ai/reference/api/): the /v1 REST surface + wire types.
- [OpenAPI spec](https://docs.collimate.ai/openapi.yaml): machine-readable API description.
- [Errors](https://docs.collimate.ai/reference/errors/): stable machine `code`s and retry guidance.
- [Limits](https://docs.collimate.ai/reference/limits/): per-tier concurrency, session, and rate limits.

## Operate

- [Self-hosting (Helm)](https://docs.collimate.ai/self-hosting/helm/): run the whole control plane in your own Kubernetes.
- [Security](https://docs.collimate.ai/self-hosting/security/): KVM isolation, per-fork egress, tenant isolation, metadata block.

## Pricing

- [Plans & pricing](https://docs.collimate.ai/pricing/): Demo / Pro / Enterprise; suspend stops the meter.