puras

Agentic backend-as-a-service

The agentic backend
for developers.

Long-running, multi-step agents that plan, call tools, and run to a finished result. Server-side. One API key. Any app.

Live · No signup

See it run.

Open a playground, send inputs, watch a real agent work end to end.

How it works

Prompt to production.
Three steps.

No servers. No queues. Write the agent, push it, call it.

  1. 01

    Write a skill

    A system prompt plus the tools your agent can call. The model plans, uses them, and iterates until the work is done.

  2. 02

    Push your skillpack

    Bundle it and deploy as one versioned release. Activating a new version is a rolling switch — in-flight jobs keep running on their version.

  3. 03

    Call it as a job

    One endpoint, one body shape, from any app with a workspace API key. Submit inputs, read the result.

curl -X POST https://api.puras.co/v1/jobs \
  -H "Authorization: Bearer $PURAS_API_KEY" \
  -d '{"skill":"deep-research","inputs":{"topic":"..."}}'

One endpoint

Three ways to call it.

Pick your latency. Same job — fire-and-forget, awaited, or streamed token by token.

Async

POST /v1/jobs

Fire and forget — returns immediately, poll for the result when you want it.

Sync

?wait=true&timeout=30

Block until the job is done, up to 60 seconds. Perfect for fast skills.

Stream

?stream=true

An SSE stream of every tool call and model response, live as the agent works.

Generative media

Image, video, audio.
Built in.

Your agents call the best image, video, and audio models through one surface — billed per call, files land in the workspace drive, ready to hand back to your app.

The agent picks the model at runtime. No wiring per model.

from puras import media

img = media.run("openai/gpt-image-2", {"prompt": "a vintage red bicycle"})
clip = media.run("kuaishou/kling-v3-i2v",
                 {"image_url": img["output_url"], "prompt": "make it spin"})

Observability

Metered to the cent.

Long-running agents are opaque by default. puras makes every run auditable.

Every job has a price tag

Each job returns its total cost, with a line-item breakdown of every model call — which model, how many tokens, what it cost.

See the agent think

A full event log per job, streamable live — every tool call, model response, and error as it happens.

Pay only for what runs

Pay-as-you-go, no subscriptions or minimums. Top up your workspace balance once; jobs draw it down as they run.

Dev loop

Deploy without leaving
your editor.

Push your whole skillpack as one deployment, one call. No git remote, no Docker, no CI. The MCP server lets Claude Code and Cursor push and tail jobs for you.

Ship your first agent
today.