Your AI assistants, inside real processes. With the right permissions.
The FlowSharp MCP server gives agents 34 tools to read, start and advance work, and nothing more than the user could do.
The MCP server is available from Optimize. Flow Agent, the built-in assistant, is available from Automate.
- For
- IT
- Partner
The problem
AI assistants become useful when they act on business systems, but giving them access means handing broad credentials to a model that can be wrong. Home-made integrations rarely have perimeters, consent, revocation and traceability. The result is an uncomfortable choice: give up the assistant, or connect it with a key that can do everything an administrator can. IT and security teams know that an overly broad access, granted today to an AI client, is tomorrow forgotten: who gave it, with which limits and until when. A serious MCP server must therefore be born with perimeters, explicit consent, expiry, immediate revocation and a readable history of what the agent did.
Chat (stylized)
OAuth consent
How it works
- 1
Create a profile or a key
The Admin defines the maximum role, allowed tools and visible flows. The perimeter is decided by whoever governs, not by the client that connects.
- 2
Connect the assistant
With OAuth, the client registers itself and the user approves a consent screen that shows exactly the role, tools and flows. Alternatively an MCP key with an expiry is used.
- 3
The agent works
It lists and reads flows and cases, starts processes, advances cases, completes tasks, uploads documents, reads analytics and map health, exports and imports maps and solutions. Each tool has its own minimum role.
- 4
People stay in command
Cases started by an agent are born awaiting confirmation; the history shows the AI client and the person who owns it. Whoever checks can always see who, on behalf of whom, did what.
Specifications
| Tools | 34, each with a minimum role and a read-only / destructive annotation |
|---|---|
| Areas covered | flows, cases, tasks, documents, events, dashboards and analytics, Flow Agent, Process Map, Solutions |
| Transport | Streamable HTTP (legacy SSE supported) |
| Authentication | OAuth with dynamic registration and PKCE, or MCP key |
| Perimeter | effective access = minimum of profile and the user's current permissions |
| Keys | tool and flow whitelist, expiry (90 days preset), never in the query string |
| Reliability | schemas enforced before every change; safe retries with expected state and idempotency |
| Limits | 120 requests per minute per IP and per key |
An operational example: from consent to case
A user connects their assistant to the MCP server. The client registers and a consent screen appears with the role, tools and flows listed; the user chooses Approve or Deny. With consent given, in a chat the user writes: start a purchase request for 3 monitors. The assistant uses the authorized tools and the case is created, but it stays awaiting confirmation, because a start by an agent requires a person. Effective access is the minimum between the chosen profile and the user's current permissions: if the user's role is reduced, the agent loses power immediately. The diagram beside it is a stylized chat with the OAuth consent, not an official screenshot. The server exposes 34 tools, each with its own minimum role.
How to verify the agent's access and actions
For an IT manager the verification has three parts. First, registered clients and keys: each key has a tool and flow whitelist and an expiry, preset to 90 days, and authorizations can be revoked at any time. Second, the perimeter: compare the tool the agent can use with the minimum role required and check that visible flows are only the intended ones, remembering that per-folder permissions apply to the agent too. Third, the actions: the case history shows the AI client and the owning person, and cases started by an agent are born awaiting confirmation. The limits of 120 requests per minute per IP and per key contain unusual use. With these elements you can show a reviewer that the agent never had more power than the person who authorized it.
Related features
- OAuth connection for ChatGPT and claude.aiProfile created by the Admin, automatic client registration and consent showing role, tools and flows.
- Scoped MCP keysWhitelist of tools and flows, expiry preset at 90 days, never in the query string.
- Start confirmation for AI and agentsA case created by an AI assistant is born waiting: it runs only after a person presses Confirm start.
- Agent attributionHistory shows the AI client and the person holding the key; the cost goes to the key.
- Built-in Flow AgentA conversational assistant always at hand (⌘/Ctrl+J), always within the user's permissions.
- Guided import of maps and solutions by agentsPreview by default; applying requires a request identifier.
Use cases
Frequently asked questions
Which assistants are compatible?
Clients that support MCP, such as ChatGPT and claude.ai, and agents built with compatible frameworks.
Can an agent see all processes?
No: it sees only the flows and tools of its profile or key, within the user's permissions.
Can I revoke access?
Yes, immediately, from the user profile or from the management of registered clients.
What happens if the agent retries a request?
Operations support idempotency and expected state: no duplicates and no wrong transitions.