Reach a whole catalogue of models through one endpoint your existing SDK already understands. Put an agent runtime behind it, or hand that agent a complete Linux desktop. Routing, scaling and uptime are ours to worry about; you pay for what you actually run.
Staking and credit units settle on Robinhood Chain. How staking works
Models behind one key
Managed agent runtimes
Workspace regions
Endpoint to integrate
Routes to every major vendor
What you can build
Use it as a plain model gateway, as a runtime that keeps an agent online, or as a whole computer that agent operates. Same account, same balance, same dashboard.
Model API
Point the SDK you already use at our base URL and reach every model on the list with the key you already hold. Routing picks a healthy provider; failures fall through instead of surfacing.
Agent runtime
Pick a runtime, give it a persona and a model, connect the places it should speak, and deploy. There is no container to build and no machine to keep alive.
Linux workspace
A dedicated machine running a keyboard-first Linux desktop, streamed to a browser tab. Choose the runtime that lives on it and let long jobs keep going after you close the lid.
Model API
The request body is the one your client library already builds. Swap the base URL and the key, keep everything else, and switch models by changing a single string.
curl https://api.agentplatform.store/v1/chat/completions \
-H "Authorization: Bearer ap_sk_..." \
-H "Content-Type: application/json" \
-d '{
"model": "atlas-5.6-solo",
"messages": [
{ "role": "user", "content": "Summarise this changelog." }
]
}'200 OK · atlas-5.6-solo · routed in 0.4s
Rates in the open
Input and output are metered per million tokens, at the rate shown on the row. No markup is added on the way out.
Sample data — illustrative figures, not live numbers.
| Model | Type | Input / 1M | Output / 1M | Routes |
|---|---|---|---|---|
atlas-6-wide Northwind | Multimodal | $0.55 | $2.76 | 9 |
lumen-5-pro Calder | Multimodal | $1.32 | $6.62 | 7 |
swift-3.8-flash Meridian | Multimodal | $0.08 | $0.41 | 9 |
atlas-5.6-solo Northwind | Multimodal | $0.28 | $1.65 | 10 |
delta-v4-flash Kestrel | Chat | $0.33 | $0.99 | 3 |
flare-2.5-image Northwind | Image | $1.65 | $9.90 | 5 |
Live on Robinhood Chain
Stake $AGENT for a receipt token, lock the receipt to issue credit units, then stake those units for a daily allowance of model credit and a discount on eligible usage.
Each staked unit carries a fixed daily allowance. The allowance resets at midnight UTC and does not accumulate; unstaking stops it from the following UTC day.
AGENT
Staking mints a receipt token one for one. The receipt can earn funded rewards and can be locked as collateral to issue credit units. Unwinding the position back to AGENT takes seven days.
Credit unit
Credit units are transferable. An active stake draws a fixed daily allowance of model credit per unit, plus a small discount on eligible usage above the cost floor. Unwinding a unit takes a day.
A credit unit is burned when the collateral behind it is redeemed, which releases the locked receipt. Buyback is not switched on in this release. Read the mechanics.
Platform
Deploying is the easy half. The rest is uptime, memory, visibility and the ability to change your mind afterwards.
Attach an agent to the chat surfaces your users already open, or leave it headless behind an endpoint of your own.
The runtime keeps the thread, the persona and the long-term notes, so a reply three days later still knows what happened.
Status, transcripts, spend and restarts sit on a single screen instead of scattered across logs and terminals.
Health checks run continuously and rebuild a failed container by themselves. Nothing waits for you to notice.
Agent lifecycle
On a machine of your own, stopping an agent means losing it. Here the platform holds the memory, so you can park one while it is idle and bring it back exactly where it stopped.
Sample data — a scripted console transcript, not a live runtime.
Compute
$0.00/hr
billing paused
Wake to warm
~15s
persona and notes restored
Workspaces
A workspace is a whole Linux machine — browser, terminal, file system — with one runtime living on it. Tasks keep going on the server even when your own laptop is shut.
Sample data — the terminal below is a mock-up, not a live session.
orders.internal/exports
| id | status | amount |
|---|---|---|
| 2141 | refund | 42.00 |
| 2142 | paid | 128.50 |
| 2143 | paid | 86.00 |
| 2144 | refund | 19.20 |
| 2145 | paid | 240.00 |
workspace-tokyo-01 · atlas-5.6-solo
you pull this week's orders and reconcile the refunds
agent on it — exporting from the back office, then running the script
browser → orders.internal/exports
shell → reconcile.py --week 35
214 rows reconciled → ~/reports/week-35.csv
$
A keyboard-first tiling desktop with a browser, an editor and the usual command-line tooling already installed.
The screen streams into a browser tab. Look in whenever you like, or grab the keyboard and finish the job yourself.
Work continues after you close the tab. Nine regions to place it in, from western Europe through to Tokyo.
From $15 a month for the machine and the runtime bundled with it. Model tokens are metered separately.
Connections
How it goes
Choose where it should speak, write the persona it should hold, and pick the model it thinks with.
Attach chat channels, webhooks and credentials. Secrets are injected at start and never printed into a log.
It stays online, answers on its own, and reports what it did. You read the panel instead of watching a terminal.
Questions
If something here is still unclear, the guides go deeper and the dashboard shows the real behaviour.
Three things on one account: a gateway that fronts a large model catalogue behind a single OpenAI-shaped endpoint, a managed runtime that keeps an agent online without a server of your own, and an optional Linux workspace that agent can drive like a person would.
The catalogue spans several vendors across chat, multimodal, image, video and embedding work. You pick a default model at setup and override it per request; the same key reaches all of them.
Creating an agent is a short wizard, and provisioning depends on the runtime and the region you chose. Progress shows in the dashboard. A fresh Linux workspace takes noticeably longer than an agent landing on an existing worker.
Each runtime gets its own isolated container and encrypted volume. Prompts pass through memory to be routed and metered, and are not kept or used for training. Secrets are encrypted at rest and injected at start, never printed into logs.
Yes. Runtimes can call outward to your services over webhooks and plain HTTP, and connections let you attach credentials without pasting them into a prompt.
Health checks run continuously and a failed container is rebuilt automatically. The dashboard keeps the event, so you can see what happened rather than guessing.
A dedicated virtual machine running a tiling Linux desktop, streamed to a browser tab. One local runtime is bundled with it. Work keeps running server-side after you close the tab, and you can take over the keyboard whenever you want.
Model calls are metered per token at the catalogue rate. Runtimes and workspaces carry a flat monthly line. One balance covers everything, and the usage log shows which request spent what.
Open the dashboard, mint a key, and send a request. Nothing to install and no machine to stand up first.