Pricing
Run it free. Pay per run in the cloud.
Self-host the open-source runtime on your own keys, free. Or deploy to puras and pay only for what each run uses — model tokens, media and tool calls, plus a flat 5%. No seats, no minimums, no subscription.
Open source
self-hosted · your own keys
Run the runtime yourself, free.
- The full agent runtime + CLI
- puras run --local — run any skill offline
- Bring your own model & provider keys
- Evals, traces & versioning, all local
- No platform fee, ever
Pay-as-you-go
no card · then usage + 5%
Deploy once, call it like an API.
- Hosted runtime — call over MCP, SDK or HTTP
- Only what each run uses — tokens, media & tool calls
- A flat 5% platform fee — no seats, no minimums
- Durable resume — long runs pick up where they left off
- Budgets, hard spend caps & per-run receipts
- Prompt caching, model pinning & publishable keys
Enterprise
volume pricing
For teams running at scale.
- Everything in Pay-as-you-go
- Volume discounts
- Priority support & SLA
- Invoicing & annual billing
- Security review & DPA
- Dedicated onboarding
Per-model token, image, video and audio rates — the base prices your usage is metered at, before the 5% — live in the model rates reference. A simple text run is a cent or two; multi-step jobs and media cost more — and every run returns an exact receipt.
Idle is $0
A deployed skill that isn't running costs nothing.
Flat 5%, all-in
The only markup — no per-call, seat or minimum fees.
Bounded by budgets
Spend caps per workspace, skillpack or skill — enforced.
A receipt per run
Tokens, model and every media call, to the cent.
FAQ
Questions, answered.
Yes. The open-source runtime runs puras run --local on your own machine and your own LLM keys — nothing to puras. It's the same engine that powers the hosted platform, not a crippled community edition, and it's open source on GitHub. Develop, test and self-host for free; you only pay once you deploy to the hosted platform and call a job.
A flat platform fee on top of hosted usage. The model rates reference lists the base prices your usage is metered at; you pay those — model tokens, media, and tool calls (a web search, say, at its usage cost) — plus 5%, billed to the cent. So $1.00 of usage is billed $1.05. That 5% is the whole markup: no per-seat, no minimum, no extra per-call fee. Self-hosting on your own keys has no fee at all.
Jobs just stop. Spend draws down your prepaid balance and pauses at $0 — there is no card on file unless you add one, so you can't be surprise-billed. The $10 credit doesn't expire, and a run that hits $0 mid-job stops without charging you for the unfinished work.
Because the work isn't the API call — it's everything around it. puras runs the agent loop, parallel steps, retries, tool calls, durable resume, evals, versioning and per-run tracing, and gives you one typed endpoint to call. You ship the skill; you skip the orchestration. The 5% is for that — not for the model access, which you could wire up yourself.
Set a budget: a spend cap per workspace, skillpack, or single skill over a day, month, rolling-30, or all-time window. enforce blocks new runs at the cap with a 402; an alert % pings you before you get there. Per-run receipts and prompt caching do the rest.
When you want volume pricing, an SLA, invoicing or annual billing, or a security review. Everything else is self-serve pay-as-you-go — reach out when you're scaling.
Your first $10 is on us.
Create a workspace, deploy a skill, call it, read the receipt. No card, no subscription — if it's not for you, you've spent nothing.