FlowSharp
AI agents and MCP

Ask the process. Flow Agent answers.

An assistant always at hand (⌘/Ctrl+J) that knows your flows, your cases and the documents you authorize.

Flow Agent is available from Automate. Those using external assistants via MCP start instead from Optimize.

  • For
  • Operator
  • Management
  • Process owner

The problem

To know "where is customer X's order" or "how do I start a purchase request" you have to know the application, the filters, the names of the flows. Occasional users get lost, and daily users waste time on repetitive searches. Knowledge of how a process works ends up with a few people, whom everybody asks, and every answer interrupts their work. A useful assistant must not know more than the user: it must know the same flows, cases and authorized documents, and answer in natural language. And it must do so without bypassing permissions or exposing reserved data to models.

Illustrative visualIllustrative values

Typed chat · ⌘/Ctrl+J

Which onboarding cases have been stalled for more than 3 days?
Case card · waiting · days waiting
Case card · waiting · days waiting
[Prepare new case] · confirmation required
Schematic illustration, not an official product screenshot.

How it works

  1. 1Open the bar

    At the bottom of the app or with ⌘/Ctrl+J. It also works full-screen and on a phone, so the question can be asked the moment it arises.

  2. 2Ask in natural language

    "Which onboarding cases have been stalled for more than 3 days?", "Explain the complaints flow to me." No filters or exact flow names are needed.

  3. 3Flow Agent searches and reads

    It searches and reads flows, cases and Library documents marked "agent context", and answers with clickable cards for cases and maps. From the answer you go straight to the object.

  4. 4It proposes a start

    It fills in the initial data and prepares the case; with "Require confirmation" on, a person confirms before it starts. The assistant prepares, the person decides.

Specifications

Accessalways-present bar, ⌘/Ctrl+J, global search
Perimeteridentical to the user's: roles, per-folder permissions, assignments
Documentsonly those marked "agent context", read page by page when needed
Chatsaved conversations, cards for cases and maps
Languagefollows the user and the organization
Availabilitycan be turned off per organization; also exposed as the MCP tool flow_agent

An operational example: an onboarding question

A manager opens the chat panel with ⌘/Ctrl+J and writes: which onboarding cases have been stalled for more than 3 days? Flow Agent searches the flows and cases the user can already see and answers with clickable cards, each with the case, the waiting state and the days waiting. From a card you go straight to the case. If a new onboarding needs to start, the assistant prepares the case with the initial data, showing a button to prepare it; with the Require confirmation option on, a person must confirm before it starts. The conversation stays saved and can be reopened. If the question concerns a Library document, the agent reads it only if it is marked as agent context. The diagram beside it is illustrative, with a stylized chat panel, not an official screenshot. Flow Agent is available from the Automate edition.

How to verify perimeter, sources and answers

With Flow Agent, verification mostly concerns what it can see and where it takes information from. The perimeter is identical to the user's: if a flow is not accessible to the asker, it does not exist for the agent, and this can be shown by asking two users with different roles the same question. Answers that cite documents report provenance, meaning document, version and page, so the reader can check the source. Because Data Shield is always applied, the history keeps only placeholders and provenance, not the reserved text read while answering. Calls appear in the AI call log with cost and attribution. If the organization prefers not to use it, Flow Agent can be turned off per organization on all channels, including the MCP tool flow_agent. For a compliance officer these are the evidence items to collect.

Related features

Use cases

Frequently asked questions

Does Flow Agent see more than I do?

No, never: it works with your own permissions.

Does it read all documents?

Only those explicitly marked as "agent context".

Can I use it from ChatGPT or claude.ai?

Yes, through the MCP server, as the flow_agent tool.

Can it be turned off?

Yes, per organization and on all channels.

Want to see it on your process?

Back to the area: AI agents and MCP