Resource hub · AI and agent builders
Authorization for teams building AI agents
Getting an agent to call a tool is easy. Deciding whether this agent, acting for this person, should take this action on this resource right now is where most teams get stuck.
The resources below cover MCP auth, tool-call checks, human approval and permission-aware retrieval. The common thread: give the agent its own identity, scope it to the task, and check every call.
Problems this page helps with
- Agents that run with a user's full access instead of a scoped, delegated one
- OAuth that authenticates the client but says nothing about the tool call
- Knowing when to pause and ask a human
The short version
Authenticating an MCP client is only the first step. Each tool call still needs a decision: is this agent, acting for this person, allowed to take this action on this resource right now?
Key concepts
OAuth on MCP
OAuth 2.1 tells the server who the client is. It does not decide whether a given tool call is allowed.
Agent identity
The agent gets its own identity, linked to the human who delegated the task, instead of borrowing the user's.
Human in the loop
Pause sensitive actions for consent or approval, based on policy rather than prompt wording.
Permission-aware RAG
Filter retrieved documents by what the user may see before they ever reach the model.
An authorized tool call, step by step
- 1
User
signs in and consents to scopes
- 2
Agent
requests a tool call
- 3
MCP Gateway
identity, policy check, approval if needed
- 4
MCP server
runs the call, outcome logged
What changes with a gateway
OAuth only
- Client is authenticated, tool calls are not checked
- Agent acts with the user's full access
- No record of which agent did what
OAuth plus per-call authorization
- Every tool call checked against policy
- Agent scoped to the delegated task
- Each call logged with human, agent, tool and decision
How Permit handles it
- 01
Put a gateway in front of MCP servers
Permit MCP Gateway handles OAuth, consent and per-call policy checks without changes to your MCP servers.
- 02
Scope agents to the task
Delegation is explicit, so a subagent can never hold more access than the agent or human that started it.
- 03
Approvals as policy
Define which actions need a human, and the gateway pauses those calls until someone approves.
- 04
Free to try
The MCP Gateway free tier includes 1,000 MAU, 10 tenants and 7 days of audit logs.
Go deeper
MCP auth
Delegation and approval
- BlogHuman in the loop for AI agentsWhen to gate an action, which frameworks support it, and how to enforce it with policy.
- BlogWhen subagents call MCP tools, who decides?Designing delegation, token brokering and approval routing so privileges never expand.
- BlogHuman to agent delegation audit trailBinding the human, the agent, the tool and the decision into one attributable record.
Build it
- BlogAccess control in LangChainA hands-on guide to adding permission checks to LangChain agents.
- BlogRAG with fine-grained authorizationFilter retrieved documents by what the user is allowed to see before they reach the model.
- DocsMCP Gateway quickstartPut Permit MCP Gateway in front of your MCP servers and run your first authorized tool call.
- ProductPermit MCP GatewayAuthorization, consent and audit for agent tool calls.
Permit MCP Gateway has a free tier with 1,000 MAU, 10 tenants, 1 environment and 7 days of audit logs.