For humans, agents, and automations in production
Gatekeeper authenticates the caller, evaluates the requested action, and issues a scoped credential only after policy approves it.
AI usage is not just agents running on their own. It is your engineers, your analysts, and your operators driving AI against real systems. Gatekeeper records every action, the agent that requested it, and the person behind it—and turns that record into the evidence your auditors ask for.
Every AI action carries an identity chain: the person, the tool they are driving, and the system it reached. You get a live picture of AI adoption across the organization—which teams are using it, against which systems, and how often. An agent action tied to a named employee is not a log line. It is an access record, and your existing access review process already knows how to consume it.
Gatekeeper maintains an inventory of every AI agent and non-human identity operating against your systems, with an owner and an authorized scope for each. Permission is granted at the individual tool call, not the service and not the session. Your agent can comment on the pull request. It cannot merge it.
Each request produces a decision receipt: person, identity, action, policy, decision, and reason. Allowed and denied are both recorded. Filter and export by person, team, system, or date range when an auditor, a customer security questionnaire, or an incident review asks.
Gatekeeper sits between the caller and the system. It verifies identity, evaluates deterministic policy, releases a scoped credential, and records the result.
A human, automation, or AI tool authenticates through the gateway. Identity is verified before any request proceeds.
The policy engine checks who is asking, what they want, when, and under which conditions.
On approval, the vault releases the credential directly into the session, scoped and time bound.
The request, decision, approval, and secret use are written to an append-only exportable log.
Agents inherit the access of whoever runs them. Gatekeeper replaces that broad permission with a decision at the individual tool call. Not the service. Not the session. The call.
No matching approval quorum. The request did not execute.
Gatekeeper sits in front of the services your humans and agents use every day. Connect a native MCP server, or wrap any REST or gRPC API behind the same policy engine and audit trail.
Microsoft 365 Copilot
When a Microsoft 365 Copilot agent calls an MCP server or API, Gatekeeper verifies identity, evaluates the exact action, and releases a scoped credential only after the request passes policy or a human approves it.
Routing model traffic and enforcing what a tool call can touch are different problems. Gatekeeper comes from protecting assets where a single mistake is irreversible, and the difference shows up in the architecture.
Architecture and controls are published for the people who approve the deployment. Legal terms are shared with qualified teams during review.
Sample attestation reports and full certification documents are available at trust.station70.com.
Visit trust centerThe roadmap extends Gatekeeper's policy model to REST, gRPC, and SDK traffic. Planned benefits include:
Reject disallowed work before downstream services consume compute, API quota, or spend.
Inspect sensitive data, secrets, outputs, and destinations before a request leaves the boundary.
Stream identity, action, policy, decision, credential, and latency fields into the security stack.
Move authorization and secret-handling context out of prompts and into deterministic policy.
Monitor mode answers the exposure question before a single policy is enforced. Forward-deployed security engineers help scope the first workflow and run the pilot with your team. You do not hire for this.
Point your MCP clients at Gatekeeper, drop the SDK into your automation code, or register connectors from the admin console. Your team keeps the clients they already use.
See every tool call your people and AI tools are already making against your own systems, including the ones policy would have denied, with nothing enforced and nothing to break.
Start from prebuilt policy templates, then refine them in the console or as code through the API. Deny by default from the first enforced request.
No new headcount required: policy is code your existing team already knows how to review, templates cover the common cases, and the managed options carry the operational load for you.
Anything else, get in touch at info@station70.com
Gatekeeper is a policy gateway between your people, your AI tools, and the systems they reach. Every request for a credential, tool, or action is checked against deterministic policy you control before it executes, at the level of the individual tool call. It includes an MCP gateway, a policy engine, and a zero knowledge secrets vault in one product.
LLM gateways govern what AI says: they route model traffic, manage prompts, and optimize latency and cost. Gatekeeper governs what AI does: which real systems a tool call can reach, with what credential, under which conditions. The two solve different problems, and many teams run both.
No. Secrets are encrypted end to end and released directly into a session under policy. The gateway itself runs inside a secure enclave, so there is no server outside it holding a key to your secrets. You can verify the enclave attestation yourself.
No. Policy is deterministic rules, not another model making a judgment call, so there is no prompt to inject and no classifier to trick. A request either satisfies the rules you wrote or it is denied before it executes.
Yes. For sensitive actions you can require a human in the loop. Gatekeeper routes the request for a fast, one click approval before any secret is released or action is taken.
Gatekeeper fails closed: an outage denies tool calls until it recovers. The blast radius of a Gatekeeper failure is a denied request, never an unauthorized action against your systems.
No. You start in monitor mode with nothing to enforce, move to prebuilt policy templates, and manage policy as code your existing team already knows how to review. SaaS, managed dedicated instance, or self hosted: none of them require a new hire to maintain.
Every request, policy decision, approval, and secret use across humans and agents is written to an append-only log you can export or stream to your SIEM and eDiscovery platforms for security review and compliance.
Yes. Put the MCP servers and API-backed actions used by Microsoft 365 Copilot agents behind Gatekeeper. Gatekeeper applies deterministic policy, scoped credentials, human approval, and append-only evidence to each request before it reaches the target system.
Our forward-deployed security engineers connect one workflow with you, run it in monitor mode, and map the allowed, denied, and approval paths. Nothing is enforced until your team is ready.
