
MCP servers One front door for every tool
Argent Strategic Intelligence answers the Model Context Protocol itself, so every tool call your agents make is checked against a permission, refused by default, and recorded in a signed ledger.
Agents now call tools, and tools run code
The Model Context Protocol (MCP) is how AI agents reach tools: a server offers tools, and an agent calls them to query a database, read files or open a ticket.
The MCP specification itself says tools "represent arbitrary code execution and must be treated with appropriate caution", and that MCP "cannot enforce these security principles at the protocol level" (MCP specification). Someone has to.
Argent Strategic Intelligence answers MCP itself, at the gateway's /mcp endpoint. Agents connect to Argent; Argent connects to your servers. Every tool call is checked against a permission first, and nothing is allowed by default.
One governed front door for every MCP server your agents use.
What Argent checks on every call
| Check | What happens |
|---|---|
| Permission | Each call needs a permission naming the server and the tool, or the whole server, for the caller's own tenant. No permission, no call. |
| Off and deny rules | An admin can switch a server or tool off for the tenant, a unit or a person, or deny it by name. These rules are checked first, and a rule list that cannot be read refuses. |
| Listed equals callable | In Enforce, an agent's tool list shows only the tools it could actually call. |
| Arguments | A permission can be narrowed to the arguments: SELECT-only for a SQL tool, or writes only under one folder. Raw argument values never enter the signed record. |
| Poisoned arguments | Shell-injection and path-traversal shapes are refused before they reach the server, the class of flaw behind CVE-2025-53109 / CVE-2025-53110. |
| New or changed tools | With pinning on Enforce, a tool that is new or whose definition changed after approval is refused and left off the list until an admin approves it. |
| Private data heading out | A call to an outside-facing tool right after a read of private data in the same session is flagged, or blocked when you set it to Enforce. |
| Tool results | What a tool sends back is marked as untrusted text before the model reads it. An optional AI scan can score it first. |
How it connects
You register each server once; agents never talk to it directly.
- Any common transport. Streamable HTTP, the older HTTP with server-sent events, and local stdio servers that Argent starts as a subprocess on first use and reuses.
- The server's credential stays in your vault. A server that needs a key or token gets it from the vault at call time; if it cannot be resolved, the call is refused before anything is sent. Agents never hold it. Works with HashiCorp Vault and OpenBao, AWS Secrets Manager, CyberArk Conjur or Infisical.
- Sessions tied to a person's agent. Every follow-up call in an MCP session is resolved again to the agent, its tenant and the server it is bound to.
- Standard sign-in for MCP clients. Argent runs the standard OAuth sign-in for MCP clients, including sign-in from devices without a browser. Access lasts one hour and the renewal token changes on every use. If a used code or token is ever replayed, Argent cancels the whole sign-in and records it in the signed ledger.
- Enterprise sign-in without a prompt. An Okta Cross App Access assertion (ID-JAG) can be redeemed for MCP access, only from the Okta issuers trusted for that tenant.
- Tenant fenced. One tenant's server settings can never be used to reach another tenant's servers.
What you configure
An admin sets this up once per tenant, then grants tools as teams ask for them.
- Add your servers. Register them by hand (transport, address or command, and the vault reference for any credential), or import them from the Official MCP Registry, the Docker MCP Catalog or your own registry. Imported servers arrive switched off.
- Approve the tools. Argent reads each server's tool list and holds every new tool under Needs a decision until an admin approves or denies it.
- Grant access. On the MCP Gate page, set each server or tool to OFF, BY POLICY or EVERYONE, then name the roles, groups, agents and people who are allowed or denied.
- Point the clients at Argent. The console generates copyable settings for Claude Code managed settings and for VS Code that keep the client on the gateway.
- Watch, then enforce. In watch mode a call that would be refused still goes through, and the would-be refusal is signed, so you can tune before anything breaks. See watch before you enforce.
Admins and operators can also request a server or access to a tool from the console; the request can open a ticket in ServiceNow or Jira Service Management.
Every decision is on the signed record
Every allowed and refused tool call is a signed row in Argent's audit ledger, with the server, the tool and the reason.
- The MCP Gate page shows every server and tool with its state, who holds it by name, and the last 24 hours of calls and refusals.
- "Who may call this tool?" is answered from the same rules the gateway enforces.
- Argent works out which permissions each agent actually used, day by day, so unused ones can be trimmed.
- Tools an agent declares in its model requests that belong to no registered server are listed too, so tools that skip the gate still show up. See shadow AI discovery.
Hand the record to an auditor with one-click audit evidence.
Where this fits
- Capabilities: agent reach control (the film and data sheet), prompt injection defense (tool results marked as untrusted), keyless agents (server credentials stay in the vault), watch before you enforce.
- Goals: threat defense, agent identity and ownership, audit and accountability.
- Use cases: controlling agent tool access, securing AI coding agents.
- Coding agents Argent governs: Claude Code, Codex CLI, Cursor, GitHub Copilot in VS Code.
- Also in MCP and agent-to-agent: Official MCP Registry, Docker MCP Catalog, the internal MCP registry, and A2A agent-to-agent.
Browse all integrations.
Want this running on your own use cases?
Book a demo