AI & Agents (overview)

Crevio is AI-native: anything a user can do in the dashboard, an agent can do through the API. The same controllers that power crevio.co/app power the REST API, the TypeScript SDK, and the MCP server — so an AI agent has full parity with a human operator. It can create products, run checkouts, send broadcasts, post to social, build and deploy a website, place phone calls, and research the web. Crevio is the platform that runs your business; the agent surface is how you let AI run it programmatically.

The agentic model

There are three layers, from “tell the agent what to do right now” to “let the agent run your business on its own.”

MCP / agent surface

Delegate work to the Crevio agent from any MCP client (ask_crevio, runs, send_message), or let the connected LLM discover and call the entire Crevio API in code (api_search + api_execute).

Tasks

A persisted instruction + schedule + agent + autonomy level. Crevio’s own AI workforce, running on triggers you define.

Events

Internal and third-party signals (an order paid, a form submitted) that fire Tasks automatically.

”Anything a user can do, an agent can do”

This is the core design principle. There is no separate, reduced “agent API.” When an agent calls POST /v1/products, it routes through the exact same controller, authorization, validation, and tenant isolation as a human clicking Create product in the dashboard. Two consequences:

Same capabilities. Every domain in the API catalog — products, checkouts, customers, discounts, email, socials, sites, domains, calls, images, web research — is available to an agent.
Same rules. The same gotchas apply: params are unwrapped (top-level), associations are bare prefix-id strings (product: "prod_abc123"), money is in cents, and creation order matters (a product needs a price variant before it can go active).

The MCP / agent surface

The MCP server has two surfaces. Delegation hands the Crevio agent a job in plain English — ask_crevio waits for the result, start_chat / wait_for_run / send_message follow and continue longer jobs — and every delegated job is a Task with one run. Direct API access is for when the connected LLM wants the controls itself: rather than exposing one MCP tool per endpoint (which burns tens of thousands of tokens before the agent does anything), Crevio uses Code Mode, and the agent writes Ruby code against two tools.

api_search
api_execute

A read-only tool to discover endpoints. The agent writes Ruby to filter a tools array describing every operation.

tools.select { |t| t[:path].include?("discounts") }
     .map { |t| "#{t[:method]} #{t[:path]} — #{t[:summary]}" }

A tool to run operations — get, post, patch, delete — and chain many calls in a single round-trip.

product = post("/products", name: "My Course", status: "draft")
post("/price_variants",
  product: product["id"],
  name: "Standard", amount_type: "fixed",
  amount: 4900, currency: "usd", billing_type: "one_time")

LLMs write code better than they navigate large JSON schemas. The agent discovers only what it needs, keeps its context lean, and a multi-step workflow that would otherwise take 10+ tool calls collapses into one execution. See the MCP guide for the full pattern, security model, and connection setup.

The two agent types

Every Task (and the dashboard chat) runs as one of two agents. Pick the one whose job matches the work.

Agent	What it does	Typical work
`business` (default)	Operates your business through the API — the marketing/ops/sales surface.	Launch a product, write and send broadcasts, post to social, research competitors, follow up with leads, place calls.
`engineering`	Writes and ships code on one of your Sites. The site’s repo is mounted as its workspace; it can build and deploy.	Add a page to your website, fix a bug, redesign a section, ship a new feature, deploy to production.

The engineering agent is scoped to a specific site via site_id (the repo is mounted as its workspace and defaults to your account’s default site). The business agent works across your whole account.

Tasks: the agent as an autonomous workforce

A single API call does one thing once. A Task is a standing instruction: a prompt, an agent, a schedule, and an autonomy level, persisted on your account. Crevio runs it on the trigger you choose — once, on an interval, on a cron schedule, or whenever an event fires — and each execution is a Task Run you can inspect, approve, or respond to.

curl -X POST https://api.crevio.co/v1/tasks \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Weekly sales digest",
    "prompt": "Summarize this week’s orders, revenue, and top products, then email me the report.",
    "trigger_type": "cron",
    "cron_expression": "0 9 * * MON",
    "timezone": "America/New_York",
    "agent": "business",
    "approval_mode": "autonomous",
    "delivery_methods": ["email", "notification"]
  }'

This is what makes Crevio an autonomous workforce rather than just an API: you describe outcomes once, choose how much oversight you want, and the agent does the work on schedule — surfacing results (or asking for approval) through the delivery channels you pick.

Task runs consume credits — every AI operation (messages, media generation, web research, calls) is metered pay-per-call. Check your balance and spend with GET /usage.

The building blocks

Primitive	Prefix	What it is
MCP tools	—	`ask_crevio` & co. delegate work to the agent; `api_search` + `api_execute` discover and call the whole API in code.
Task	`task_`	A standing instruction: prompt + agent + trigger + autonomy + delivery.
Task Run	`trun_`	One execution of a Task — status, summary, credits consumed, tool calls.
Event	—	An internal signal (`order.paid`, `form_submission.created`) that can trigger an event-Task.
Event Source	—	A third-party trigger (a connected app’s event) that fires an event-Task.
Site	`site_`	A website the `engineering` agent builds and deploys to.

Next steps

Tasks (your AI workforce)

Triggers, agents, autonomy modes, delivery, and responding to supervised runs.

Events & Event Sources

Fire Tasks automatically from internal events and third-party apps.

MCP Server

Connect Claude, Cursor, or any agent to the full API via Code Mode.

Connections

Run a connected app’s tools, call its API directly, and gate every write behind an approval.

Usage & credits

See how AI operations are metered and check your balance.

Getting started

AI & agents

Core concepts

Integrations

Recipes

AI & Agents (overview)

The agentic model

MCP / agent surface

Tasks

Events

”Anything a user can do, an agent can do”

The MCP / agent surface

The two agent types

Tasks: the agent as an autonomous workforce

The building blocks

Next steps

Tasks (your AI workforce)

Events & Event Sources

MCP Server

Connections

Usage & credits

​The agentic model

MCP / agent surface

Tasks

Events

​”Anything a user can do, an agent can do”

​The MCP / agent surface

​The two agent types

​Tasks: the agent as an autonomous workforce

​The building blocks

​Next steps

Tasks (your AI workforce)

Events & Event Sources

MCP Server

Connections

Usage & credits

The agentic model

”Anything a user can do, an agent can do”

The MCP / agent surface

The two agent types

Tasks: the agent as an autonomous workforce

The building blocks

Next steps