Skip to content

Authorization

An MCP call is authorized in layers. Each layer can only narrow what the previous one allowed, and a refusal at any layer ends the call.

Layer Question Refusal
Origin check Is the browser origin allowed? 403 origin_not_allowed
Verified caller Did the capability gate admit a verified caller, and does the profile accept its credential kind? 401 unauthorized
Scopes Does the caller hold every scope the tool requires? 403 insufficient_scope
IMcpToolAuthorizer Does every registered authorizer allow this tool for this caller? 403 forbidden
The operation Does the command or query pass its own authorization and validation? Tool error, isAuthorized: false or isValid: false

The full admission order is in the options reference.

Arc MCP never reads HttpContext.User or a header to decide who is calling. The capability gate verifies the credential and stores a verified caller for the request. MCP reads only that. If the gate admitted no caller, no MCP method runs and the response is 401.

The verified caller carries the person, the tenant, the scopes of the credential and the credential kind. The command or query that a tool runs sees the same identity and tenant it would see for a REST call.

Kind How it arrives Accepted by profile
Personal access token (PAT) Authorization: Bearer <token>, verified against your IPatAuthority. PatOnly (default), Combined
OAuth grant Authorization: Bearer <token> from your authorization server, verified by your installation. OAuthOnly, Combined
Browser session (signed forwarded identity v2) A signed envelope forwarded by a trusted proxy. None. MCP refuses it with 401.

A signed forwarded identity is how a signed-in person reaches the application through the proxy. Under the PAT installation that session is what lets the person issue a PAT; the agent then calls MCP with the PAT, not with the session. Selected capability endpoints, MCP among them, refuse forwarded envelopes. A PAT that was forwarded through the proxy is still verified against your PAT authority.

Arc does not issue PATs through an endpoint of its own and does not issue OAuth tokens. See Connect clients for the profiles.

A PAT is bound to one person and carries a fixed set of scopes. It can hold only scopes on PatOptions.AllowedScopes, never a scope with an admin or approve segment, and it lives at most 90 days (PatOptions.MaximumLifetime). The agent acts with the person’s identity, restricted to those scopes.

Scope held The agent can
None of the tool’s scopes Not see the tool in tools/list, and get 403 insufficient_scope if it calls it.
<ApplicationName>:read List and call query tools.
<ApplicationName>:write (or another write scope) Call command tools that require it. A write scope does not grant read.
The above plus a tool’s [RequiresScope] extras Call that tool. Every required scope must be held.

Holding the scope is necessary, not sufficient. The command’s own validation, its authorization policy and any IMcpToolAuthorizer still run, so a person’s PAT can never do more than the person could do through the application.

There is no MCP session: the transport is stateless and each request is verified again, so a revoked token stops working on its next call.

  • Add required scopes with [RequiresScope("...")]; see Scopes and descriptions.
  • Implement IMcpToolAuthorizer for rules a scope cannot express, such as “only members of this organisation”. See Tool sources and authorizers.
  • Commands that require approval are never exposed.