Deploy an official template or your own container image. Every agent gets its own URL, private by default, and answers only callers holding one of its access keys. Bring your own model key from OpenAI, Anthropic or Google.
No cluster to run, no ingress to configure, no certificates to renew.
Start from the customer support or code reviewer template, a blank agent with your own prompt, or any public container image that serves /health and /invoke.
Each agent answers at https://<name>.run.agntspark.com, with its certificate issued and renewed for you.
New agents only answer callers holding one of their access keys. Give a key to an app, revoke it when you're done, and cap how many requests one caller may send per minute.
Your OpenAI, Anthropic or Google key is encrypted at rest and handed only to your agent's containers. Model usage stays on your provider bill.
Run several replicas of an agent, scale by hand or automatically within your plan, and see live CPU and memory, logs, and metered usage by the hour.
Everything the console does is a REST call. The Python SDK and agntspark command line cover deploys, scaling, logs, access keys and calling your agents.
AgntSpark runs on a single host today. That's enough for the alpha, and it's what the platform actually is.
console + API agent URLs
agntapi.agntspark.com *.run.agntspark.com
│ │
▼ ▼
┌───────────────────────────────────────────┐
│ Caddy — TLS, routing │
│ │ │ │
│ ▼ │ asks: who is │
│ ┌───────────┐ │ this agent? │
│ │ Gateway │◀───────────┘ │
│ │ auth, │──▶ Postgres (agents, keys, │
│ │ quotas, │ usage) │
│ │ Docker │──▶ Redis │
│ └─────┬─────┘ │
│ │ starts / stops │
│ ┌─────▼─────┐ ┌───────────┐ │
│ │ agent A │ │ agent B │ … │
│ │ container │ │ container │ │
│ └─────┬─────┘ └─────┬─────┘ │
└────────┼──────────────┼───────────────────┘
▼ ▼
your model provider and your own APIs
Requests to an agent go from Caddy straight to one of its containers; the gateway only says where to send them, checks the access key, applies rate limits and counts the request. Agents can't reach each other, the database, or the cloud metadata service, and run without extra privileges. They do share one kernel, which is a reason the alpha is invite-only.
Hosting only. You pay your model provider directly for what your agents think.
The alpha is small on purpose. Tell us what you'd build and we'll send you a code.