Apify

Reach extends well beyond the web. The integrations catalog covers automation platforms (Zapier, Make, n8n), vector databases and LLM frameworks. Through MCP connectors, Actors can call third party MCP servers such as Notion or Slack on a user's behalf, and webhooks POST run events to any URL. The MCP server itself exposes Store Actors to outside agents as callable tools, many of which write to other services, so agents can read from and write to other systems through native integrations and custom tools.

Sourcedocs.apify.com/integrations/mcpread 2026-09-22

Customers chain work by wiring steps together. A webhook on a run finishing or failing can start another Actor, so a pipeline can branch on the outcome of a step. Saved tasks fix an Actor's input for reuse, Actors can hand work to other Actors in code, and request queues and retries run inside a crawl. Loops, branching, retries and fallback paths are assembled from webhooks and code rather than defined in a workflow runtime, and no versioned flow definition is documented.

Sourcedocs.apify.com/integrations/webhooksread 2026-09-22

Web data gathered by Actors lands in datasets and key-value stores, and there is no retrieval structure over the customer's own documents for an agent to query. Crawlers and vector database integrations push web content into the customer's own vector store, so the grounding happens in the customer's layer. No feature indexes, refreshes or permissions customer content for retrieval, and public web data is not the customer's corpus.

Sourceapify.com/llms.txtread 2026-09-22

One class of action waits for a person. Actors that ask for full permissions must be approved by the user before they run, and the platform returns full-permission-actor-not-approved or full-permission-actor-blocked-for-admin when that approval is missing or an admin has blocked them.

The MCP server leaves full permission Actors out entirely, because running one is a decision the user approves personally and an LLM cannot make it on the user's behalf. That consent checkpoint separates higher risk Actors from limited permission ones. No approval step inside a run, escalation rule or routing of approvals into other channels is documented.

Sourcedocs.apify.com/integrations/mcpread 2026-09-22

SOC 2 Type II compliance after an independent audit is stated in Apify's security documentation, with the report available on request through its Trust Center at trust.apify.com. Customer facing controls sit beside it. Organization accounts assign every member a role with configurable permissions per resource type (Actors, tasks, storage), organizations can require two factor authentication and cap session lifespan, and the shared responsibility page lists SSO and scoped API tokens among the controls Apify provides.

Sourcedocs.apify.com/securityread 2026-09-22

Every Actor run keeps a log of what it did, a run record with status, statistics and resource usage, and its input and output storages, retrievable in the Console and through the API; schedules keep their own log. Native monitoring charts run statuses over the last 30 days and metrics over the last 200 runs of any Actor or task, and alerts fire during or after a run when a metric crosses a threshold or a run ends in an unexpected status. Runs and their storages are retained for a period set by plan. No audit log separate from run logs and no export to a SIEM is documented.

Sourcedocs.apify.com/actors/running/monitoringread 2026-09-22

Run state persists for the Actor's code to read. Actor.useState saves an Actor's progress so a run resumes where it left off after a server migration, named key-value stores, datasets and request queues keep data across runs, and default storages are kept for the plan's retention period. The Actor reads that state to decide what to do next, so it is state rather than a setting. These are storage primitives the developer wires up, with no memory layer of stated types.

Sourcedocs.apify.com/actors/development/builds-and-runs/state-persistenceread 2026-09-22

The platform, with its scheduling, storage, proxies and monitoring, is a hosted cloud service, and no region choice, dedicated cloud or self hosted platform is documented. Actors can also run locally, and Apify's open source Crawlee library runs on the customer's own infrastructure, though without the platform's services.

Sourceapify.com/llms.txtread 2026-09-22

A browsable catalog, Apify Store, holds ready to run Actors, 73,051 on the homepage on 22 September 2026, that scrape named sites, generate leads, monitor social media and track competitors. Each has its own README, input form and pricing and does its job when selected, and most are built by independent creators. Apify says an Actor is a serverless program rather than an AI agent, but each one is a complete working job a buyer adopts rather than a connector. Actor templates for Python, JavaScript and TypeScript add starting points for builders.

Sourceapify.comread 2026-09-22

Work starts without a person initiating it. Schedules run Actors and saved tasks on a recurring timetable, with a schedule log, and webhooks fire on system events such as a run finishing or failing, for example to start another Actor. Runs can also start from the API, the MCP server and integrations such as Zapier, Make and n8n. No chat or email channel is documented as a starting point.

Sourcedocs.apify.com/integrations/webhooksread 2026-09-22

Apify serves tools and data to whatever model the customer's agent runs. Any MCP client (Claude, ChatGPT, Cursor and others) can connect to its MCP server, its docs cover frameworks such as CrewAI and Agno with a note that other LLM providers can be swapped in, and AI Actors take the customer's own model keys. Only Apify AI in the Console chooses its own model; for the customer's agent, model choice stays with the customer.

Sourceapify.com/llms.txtread 2026-09-22

A documented REST API covers Actors, runs, tasks, schedules, webhooks, storages and the Store, alongside the apify-client packages for JavaScript and Python, SDKs for building Actors, a CLI, and a hosted MCP server with OAuth or local stdio. An RFC 9727 API catalog and published Agent Skills describe the surface for agents, and developers can package, version and publish their own Actors.

Sourcedocs.apify.com/llms.txtread 2026-09-22

Output gets a quality gate. A dataset schema validates each item an Actor writes at the field level, and monitoring alerts on dataset field statistics, for example when a field stops being filled, alongside alerts on run metrics and status. That makes quality gates configurable for output data. No test harness for fixtures or datasets before production and no scored evaluation of output quality over time is documented.

Sourcedocs.apify.com/actors/running/monitoringread 2026-09-22

Apify Actors drive real browsers on its cloud. The SDK's Playwright and Puppeteer crawlers run browser sessions with persistent session and cookie management, and the documentation shows Actors acting on a page by clicking, typing and pressing keys, behind Apify's anti blocking and proxy services. That control of a real interface runs remotely on the vendor's infrastructure.

Sourcedocs.apify.com/llms.txtread 2026-09-22