Everything buyers and builders ask before putting AgentGate Labs in the agent path

AgentGate Labs is a runtime authorization layer for AI agents and MCP tool calls. It sits between agents and the tools they want to use, then decides whether each tool call should be allowed, denied or routed to human approval before it reaches the upstream system.

AI agents are moving from suggesting work to executing work. Once an agent can call shell, database, GitHub, browser, deployment, SaaS or payment tools, a bad prompt or bad workflow can become a real action. AgentGate Labs reduces that blast radius by enforcing policy before execution.

No. AgentGate Labs does not replace Codex, Claude Code, Cursor, Cline or internal agents. It protects the tool path those agents use, especially when they connect to MCP servers or sensitive internal APIs.

MCP is the first-class integration path because many coding agents and internal assistants use it to reach tools. The same control model also applies to any agentic workflow where tool calls can be intercepted, evaluated and forwarded through a gateway.

The first buyer is usually an AI platform, security, DevOps, SRE or engineering leader whose team is already letting agents touch repositories, shells, databases, logs, deployment systems, browser sessions, SaaS APIs or customer data.

The developer keeps using their normal AI agent inside the project folder. For remote MCP servers, the agent points directly at the AgentGate Labs gateway. For local laptop tools such as filesystem, shell, local repo automation or stdio MCP servers, the agent points at AgentGate Labs Local Connector running on the same computer. The connector asks AgentGate Labs for authorization before it forwards the call to the local MCP server.

A team creates a workspace, registers an agent, copies the agent credential, registers one or more MCP server entries, then updates the local agent configuration. Remote MCP traffic goes to the hosted gateway. Local MCP traffic goes to AgentGate Labs Local Connector, which syncs tool inventory to AgentGate Labs and asks for allow, deny or approval decisions before local execution.

Not for remote MCP servers. Those can use the hosted or self-hosted gateway directly. For laptop-local tools, a lightweight AgentGate Labs Local Connector runs on the developer machine because AgentGate Labs cloud cannot directly reach local stdio processes, project files or localhost-only tools.

Desktop agents often talk to local MCP servers over stdio and those servers can access the developer filesystem, shell or repository. A cloud gateway cannot safely invoke those local processes. AgentGate Labs Local Connector becomes the MCP server the desktop agent sees, calls AgentGate Labs for the policy decision, and only then forwards allowed calls to the real local MCP server.

They should not need to change how they ask the agent to work. They only change the tool connection path. Low-risk reads can still move quickly, while risky actions such as production writes, destructive shell commands or sensitive exports are controlled.

AgentGate Labs evaluates the tool name, arguments, agent identity, environment, trust level and policy context. Depending on policy, the call is allowed, denied, or held until an approver reviews it. The upstream tool is not called until the decision allows it.

Yes. Each agent should receive its own credential so audit records can show exactly which agent requested which tool, for which environment and under which policy.

IAM authenticates identities and grants broad permissions. AgentGate Labs authorizes the live agent action. It sees the MCP tool name, arguments, target environment, risk signals and policy match before the call runs.

SIEM and logs help after something happened. AgentGate Labs is an inline control: it decides before the tool call reaches code, data, infrastructure or a SaaS API.

A tool call with no matching allow policy does not pass. This makes unknown or newly discovered capabilities safer by default, especially when new MCP tools appear.

If policy evaluation, approval lookup or required control-plane checks fail, the gateway blocks instead of silently allowing risky traffic.

AgentGate Labs does not claim to make prompts impossible to manipulate. It reduces the impact of prompt injection by preventing injected instructions from freely using powerful tools.

The product is designed to redact secret-like values in audit records and keep raw sensitive values out of logs. Detection is pattern-based, so teams should still avoid putting raw production secrets into prompts or tool arguments.

Yes. The point is not to stop agents from being useful. Policy can allow routine work, deny dangerous work and route legitimate high-risk actions to a human approval queue.

AgentGate Labs is designed for MCP-compatible clients and agents, including coding agents such as Codex, Claude Code, Cursor, Cline and internal tools that can be configured to use an MCP gateway.

The hosted gateway supports HTTP and Streamable HTTP upstream MCP servers. AgentGate Labs Local Connector supports local HTTP MCP servers and basic command-mode stdio MCP servers on the developer machine. Production-grade long-lived stdio supervision is still a known roadmap item.

Yes. For remote MCP servers, AgentGate Labs can scan tool inventory from the gateway. For local laptop MCP servers, AgentGate Labs Local Connector forwards tools/list locally and syncs the discovered tools to AgentGate Labs so policy can evaluate names, schemas, inferred capabilities and risk labels.

Yes, when those capabilities are exposed through MCP or another gatewayed tool path. AgentGate Labs can inspect and control the request before forwarding it to the real upstream system.

Yes, if browser automation is exposed as a governed tool path. This matters because browser agents can access live sessions, internal apps, customer data and workflows that create external side effects.

Yes. A user can create a workspace, register agents, connect an MCP server and validate policy decisions during the 14-day trial.

Trial users can validate the core workflow: create agent credentials, connect MCP servers, discover tools, simulate policy decisions, inspect approvals and review audit events within trial limits.

When the trial ends, payment is required to keep using AgentGate Labs for ongoing operations. The Team plan is the self-serve paid option, while Enterprise is for private deployment, compliance and custom support.

The console includes a hosted checkout flow for the Team plan. Enterprise customers use a sales-assisted commercial process because deployment, data residency, SSO and compliance needs vary.

Yes. The risk review form is for guided rollout: connect one real MCP workflow, identify risky actions, define initial policy and decide whether Team or Enterprise is the right path.

Audit records include the agent, tenant, environment, MCP server, tool, decision, risk score, matched policy, explanation, timing and redaction metadata. Approval records add approver context and decision state.

Yes. The console supports audit export workflows, including formats suitable for security reviews and operational investigation.

The approving user depends on workspace policy and team operations. In the MVP console, high-risk calls appear in an approval queue where authorized users can approve or reject them with a comment.

Approvals are designed to be scoped, expiring and replay-resistant. They are bound to the specific request context rather than acting like a broad standing permission.

Yes. The main evidence story is simple: show what agents can access, which policies control them, what was denied or approved, and how sensitive values are redacted in records.

Yes. The product includes a Docker-based deployment path and is designed for teams that need to run the gateway and control plane in their own infrastructure.

Yes. The site and console are structured around a managed workspace and self-serve Team path, with Enterprise available for private cloud or on-premises requirements.

AgentGate Labs has a production-oriented MVP with authentication, tenancy, gateway enforcement, policy simulation, approvals, audit records, billing foundation and Docker deployment. Some enterprise-grade items remain roadmap work, including broader stdio supervision, SSO/OIDC/SAML, SIEM integrations, full tracing and independent security review.

Start with one high-value MCP workflow: production database reads and writes, deployment tools, repository automation, incident-response tools, browser automation or customer-data exports. Prove allow, deny, approval and audit behavior before expanding.

For protected tool paths, the intended security posture is fail-closed. Teams should deploy AgentGate Labs with the same reliability expectations as other control-plane infrastructure.