Pipefy

Agents reach outside systems in several documented ways. Developers and IT teams connect their own MCP servers so agents query external databases in real time and act in other systems, for example creating a payment in SAP or sending a receipt through Slack after an approval. Pipefy iPaaS carries native connectors and custom API calls to external services, with credentials held in secure connection settings, and automations make HTTP requests and connect through Zapier. Pipefy's agents consume MCP as a client.

Sourcehelp.pipefy.com/en/articles/14300355-mcp-the-universal-standard-for-powering-up-your-ai-agentsread 2026-09-18

Pipefy coordinates its own agents with external ones inside a process. A flow in Pipefy iPaaS starts when a card is created, calls an external agent with the card's data, maps the response back onto the card, and hands over to Pipefy's native agents. Those agents structure the result, move the card to the right phase, such as manager approval or security review, and prepare the next step. Processes run as pipes with several phases, each able to carry its own agents and automations, so the work is coordinated across steps rather than run as a single task.

Sourcedevelopers.pipefy.com/docs/orchestrating-agents-with-pipefyread 2026-09-18

Agents answer from knowledge bases the customer maintains. A pipe admin adds knowledge base documents, plain text entries and data lookups scoped to the pipe, and AI assistants carry their own knowledge base documents, all created, updated and deleted through the API as well as the product. An agent or an individual behavior points at those sources by ID and queries them for context when it runs, so new material enters by adding it to the base rather than by retraining. Intelligent document processing reads complex documents as input to the work.

Sourcedevelopers.pipefy.com/reference/create-knowledge-base-document-mutationread 2026-09-18

A person can be made to check an agent's output before it moves forward. Inside any agent behavior, the Request human validation action sends the review to a named reviewer as a task with the phase SLA applied, including reviewers outside the company through a shared form. Every time the behavior runs, the output waits on that review before it moves on. Agents can also leave output such as a drafted email ready for a person to send, and move a card into a manager approval phase based on their verdict, so approvals sit in the workflow itself.

Sourcehelp.pipefy.com/en/articles/16005153-ai-agents-requesting-human-validationread 2026-09-18

A 2025 SOC 2 report and ISO 27001, ISO 27018, ISO 27701 and ISO 42001 certificates are published on Pipefy's trust portal, and the security page carries the same certification marks. The controls a customer uses are documented too. They include enterprise SSO with an SSO only mode that disables password sign in, two factor authentication, and role profiles that set what each user can see, edit or change. Users can be managed from an identity provider or Active Directory, and data is encrypted in transit and at rest.

Sourcepipefy.com/securityread 2026-09-18

Every agent run leaves an execution log with a tracing graph, and each node in that graph can be opened to see what the agent did at that step. Pipe admins query these logs through the API as well as in the product's tracing and logs view, so what is traced is the agent's own behavior. Separately, the platform keeps audit logs of activity, exportable as pipe level and organization wide activities reports, and AI agent usage per organization is queryable. Retention terms and a streaming export to an external monitoring system are not documented.

Sourcedevelopers.pipefy.com/reference/ai-agent-log-node-details-queryread 2026-09-18

Pipefy's AI Assistant is a chatbot that answers employee questions and creates requests through conversation, carrying conversation state within a session. Agents otherwise work on cards, and a card is the business record itself, so what an agent reads there is the application's data model rather than a memory it keeps. Knowledge bases are reference material the customer maintains. No memory layer with its own scope and lifetime is documented, so there is no stated way to review, edit, delete or scope memories by user, team or workspace.

Sourcehelp.pipefy.com/en/articles/10769512-what-is-the-ai-assistant-betaread 2026-09-18

Services and data are hosted in Oracle Cloud Infrastructure facilities in the United States, according to Pipefy's security page, and the cloud SaaS product documents no region choice for platform data. Model traffic is a different story. Under Bring Your Own Models a customer connects its own Azure OpenAI, AWS Bedrock or Vertex AI account, keeps control over the privacy, retention and residency of what the model processes, and Pipefy acts only as the orchestration interface.

Sourcehelp.pipefy.com/en/articles/13355927-pipefy-ai-governance-and-operational-architectureread 2026-09-18

Pipefy ships packaged assets a customer can adopt. Its AI Studios are packaged solutions that each do one job. They cover supplier relationship management, background checks with a consolidated trust score, procure to pay with agents working against the ERP, quotations, CRM, onboarding with data and risk validation, credit decisioning, and claims management covering validation, fraud detection and settlement. Each one stands on its own as a whole product. Plug and play process templates are included on every plan, the free Starter plan among them.

Sourcepipefy.com/pricingread 2026-09-18

Work reaches Pipefy's agents without a person asking. Each agent behavior carries a trigger event, the same events regular automations use, such as a card being created or moved to a phase. The behavior runs its instruction and actions when that event fires. Orchestrations built in Pipefy iPaaS start the same way from a card created trigger, and organization and pipe webhooks cover events leaving the platform. Event triggers are documented on the agent itself, and the AI Assistant adds a conversational channel.

Sourcedevelopers.pipefy.com/reference/create-ai-agent-mutationread 2026-09-18

The customer controls which model runs. An administrator with permission to manage AI providers sets the active LLM provider for the whole organization or for an individual assistant, choosing among system providers Pipefy manages or a custom provider configuration the organization registered. The models available for each provider can be listed, and each agent behavior can name its own provider. Pipefy states that requests on its managed path go to contracted providers such as OpenAI, Google and AWS under terms that exclude training on customer inputs and outputs.

Sourcedevelopers.pipefy.com/reference/set-active-llm-provider-mutationread 2026-09-18

The platform is callable from outside through a documented GraphQL API, and the agent layer is part of it. AI agents are created, updated, enabled, duplicated and deleted through API mutations with their full behavior configuration, agent execution logs are queryable, and LLM providers and knowledge bases are managed the same way. Service account tokens authenticate the calls, a Postman collection is published, and organization and pipe webhooks push events out. API access is included from the Business plan up.

Sourcedevelopers.pipefy.comread 2026-09-18

Changes can be tested before they reach the live process. Pipefy's automation simulation runs an automation, including an AI generation action, against a chosen sample card and returns the result, its status and any error details, so the output is previewed before the change is applied. A sandbox version opens an editable copy of a pipe, cloned from a snapshot, where phases, fields and automations are changed without touching the pipe users work in. The sample card serves as the test fixture and the sandbox as the release path.

Sourcedevelopers.pipefy.com/reference/automation-simulationread 2026-09-18

An agent behavior can switch on a web scraping capability that reads web pages, alongside web search, document OCR and calculation, so agents reach content on the web without an API. Nothing documents an agent clicking, typing or navigating in a browser or desktop application.

Sourcedevelopers.pipefy.com/reference/create-ai-agent-mutationread 2026-09-18