# Agent Next

> Agent Next builds agent-native systems for large-scale autonomous work in the cloud. Its lead product is the agent-first GPU sandbox: a GPU the way agents use a CPU sandbox, with VRAM slices from 0.5 GiB, pause and resume that keep Python state, and usage metered per busy second. Also live: Agent Next Sandbox (isolated cloud sandboxes where agents run code and tools, with an API and a browser console), Agent Next Token (an OpenAI-compatible model endpoint, access by invitation) and the open-source Agent Next Paper Trader (paper trading for AI agents on live Polymarket and Kalshi order books).

## GPU sandbox

- [GPU sandbox](https://agent-next.com/gpu/): VRAM slice of 0.5 to 28 GiB on today's RTX 5090 (`accel.vram_gb`, or tier S 1 / M 2 / L 8 GiB; XL 32 GiB needs a larger card, not live yet); a 0.5 GiB slice runs small CUDA jobs; start at 2 GiB (tier M) for training-sized allocations. Measured through the public API on 2026-10-17 on one RTX 5090 (`nvidia.5090`), the largest single torch tensor that allocated (binary search, 8 MiB steps) was: 0.5 GiB slice: 248 MiB; 1 GiB slice: 760 MiB; 2 GiB slice: 1784 MiB. About 264 MiB of each slice goes to the CUDA context and allocator.
- Pause and resume: on pause-capable templates, pausing releases the GPU and spills its state to host memory, unless other processes in your lease hold VRAM or the VRAM probe is unreadable, in which case the lease freezes, keeps its seat and bills its frozen time; resume restores your state bit-exact
- Metering: usage is metered per busy second. Read billed seconds from `meters.billed_s` on get and exec replies, and from `billed_s` in the delete reply body; create, pause and resume replies carry no meters.
- API base: https://api.sbx.agent-next.com
- Auth: send your key in the `X-API-Key` header on every request
- Create a lease: `POST /sandboxes` with `{"accel": {"vram_gb": 2}}`; run Python with `POST /sandboxes/{id}/exec` and `{"code": "..."}` (or a shell command with `{"cmd": "..."}`); then `POST /sandboxes/{id}/pause`, `POST /sandboxes/{id}/resume` and `DELETE /sandboxes/{id}`
- [sbx gpu CLI](https://github.com/agent-next/sandbox/blob/main/cli/README.md#gpu-leases): open-source `sbx` CLI verbs `sbx gpu create --vram 2`, `sbx gpu exec ID -c '...'`, `sbx gpu pause ID`, `sbx gpu resume ID`, `sbx gpu kill ID`; install it from the repo
- Remote MCP: https://api.sbx.agent-next.com/mcp/v1 (send `X-API-Key`); the tool list is `GET /v0/facts` `mcp_tools` (11 tools on 2026-10-22), including `sandbox_templates`, `sandbox_create` with a GPU template and `vram_gb`, `sandbox_exec` and `sandbox_kill`
- Plan before you create: `GET /v0/gpu/capacity` lists free seats per class and the largest slice a create is admitted for (`max_admit_gib`); `accel.wait_s` (0 to 120) makes a create wait for a seat; refusals carry a typed `error`: `gpu_capacity_exhausted` comes with a `Retry-After` header, and `gpu_class_unavailable` (400) means the class is not live
- [GPU docs](https://sbx.agent-next.com/docs/gpu/): templates and the `accel` object
- [Get an API key](https://sbx.agent-next.com/beta/): GPU is an invite-only beta: a key needs a GPU grant or a GPU-enabled tier, and the free tier has no GPU. Join the waitlist for an invite (10 GPU-hours by default on our RTX 5090 at a single site, no SLA); limits: https://sbx.agent-next.com/docs/gpu/#invite-only-beta
- [Class availability and list prices](https://agent-next.com/gpu/#pricing): a class is live or on request, and its list price applies only once it is live; the same data as [classes.json](https://agent-next.com/gpu/classes.json)
[//]: # (gpu-classes:llms:start)
- Live now: nvidia.5090
- On request, not live today: nvidia.rtx6000, nvidia.a10, nvidia.a6000, nvidia.a100-40g, nvidia.gh200, nvidia.h100-pcie, nvidia.h100-sxm, nvidia.b200
- Availability as of 2026-10-22
[//]: # (gpu-classes:llms:end)

## Sandbox docs

The full Sandbox documentation lives on the agent-infra site, not here:

- [Sandbox docs](https://sbx.agent-next.com/docs/): documentation home for the Sandbox
- [Quickstart](https://sbx.agent-next.com/docs/quickstart/): create a sandbox and run your first command
- [Limits](https://sbx.agent-next.com/docs/limits/): what a sandbox can and cannot do
- [Console](https://sbx.agent-next.com/console/): create and inspect sandboxes in the browser

## API reference for agents

- [llms.txt](https://api.sbx.agent-next.com/llms.txt): plain-text reference for the Sandbox API (E2B-compatible gateway for CPU sandboxes at https://api.sbx.agent-next.com; GPU leases use the REST, MCP and CLI verbs above, not the stock E2B SDK)
- [Agent Next docs hub](https://agent-next.com/docs/): entry point for Agent Next Sandbox and Token documentation

## Token

- [Token endpoint](https://gateway.agent-next.com/v1): OpenAI-compatible model endpoint; access is by invitation. Send the key as `Authorization: Bearer <your key>`.

## Paper Trader

- [Paper Trader](https://github.com/agent-next/polymarket-paper-trader): open-source paper trading for AI agents on live Polymarket and Kalshi order books (real prices and fees, simulated money); MCP server and CLI, `pip install polymarket-paper-trader`

## Agent Next site

- [Home](https://agent-next.com/): the GPU sandbox, products, mission and research
- [Products](https://agent-next.com/products/): the GPU sandbox, Sandbox, Paper Trader and Token
- [Status](https://agent-next.com/status/): live service status
- [Science fleet](https://agent-next.com/science/): research pages
- [Notes](https://agent-next.com/blog/): engineering and research notes