Decide whether access belongs to exact verified email addresses, exact normalized domains, or a named API integration.
Private client delivery
Client-facing AI agent with real private access
A client-facing AI agent should verify who is allowed in, keep access revocable, and leave the Agent's instructions, tools, and shared knowledge under your control.
For the usual MCP path, model-token cost to the Agent creator is $0: the customer's AI client supplies the model under that customer's subscription and limits.
Early access
What is getting in the way
Sending an unlisted URL is not an access policy. It is hard to revoke, easy to forward, and gives neither you nor the client a clear boundary around identity and authority.
What changes
Publish the Agent privately, grant exact people or domains through OAuth, or issue a scoped API key when server-to-server access is the better fit.
Private client delivery
How Prostir gets you there
List the allowed clients
Decide whether access belongs to exact verified email addresses, exact normalized domains, or a named API integration.
Choose the access mode
Use OAuth for people or create a scoped, expiring, revocable API key for an explicit machine-to-machine need.
Publish and review
Confirm the policy before publication, then monitor and revoke grants or keys without rewriting the Agent.
Early access
What you can verify
- New Agents are private by default and begin with OAuth protection rather than anonymous access.
- OAuth policy uses exact verified emails or normalized exact domains and rejects suffix-style matching.
- API keys are Agent-scoped, named, expiring, revocable, rotatable, shown once, and stored hashed.
A good fit if
- You deliver an Agent to named clients, companies, or approved integrations.
- You need the customer to use the Agent without editing its core configuration.
Not the right fit if
- You want an untracked public link to count as private access.
- You need several coworkers to edit shared Tasks, Files, Data, and Knowledge; that is a Company use case.
Questions before you decide
01Can I allow a whole client company?
Yes, with a normalized exact-domain OAuth grant when that policy is appropriate. Similar-looking domains, suffixes, and unverified email claims do not qualify.
02Can a client change the Agent?
No. A standalone Agent remains author-controlled. Clients can use their explicitly owned personal context, but they cannot mutate the shared instructions, tools, Skills, or Agent Knowledge.
03Who pays for model tokens, and what do they cost the Agent creator?
For the usual MCP path, model-token cost to the Agent creator is $0. When a customer uses the Agent through a supported MCP connector in ChatGPT or Claude, the model runs in that customer's AI environment under the customer's subscription and limits. The creator pays for the Prostir plan, not those model tokens. AI Credits are needed only when the creator separately chooses a Prostir-hosted model run, such as a supported channel or widget. BYOK is optional and bills the creator's own provider account. Paid external tools remain separate.
Explore related options
Early access
Design access before you send the link
Tell us who the clients are, how they should sign in, and what should remain owner-only. We will map the correct private Agent boundary.