
An agent can use whatever it is connected to
Every time an agent does real work, it reaches for a tool. With nothing in between, an agent can use any tool on any MCP server it is connected to.
An MCP server (Model Context Protocol) is a program that gives an AI agent tools, such as reading a file or sending a payment. One server can hold many tools, and each server is separate.
- OWASP's Top 10 for LLM Applications 2025 names excessive functionality, excessive permissions and excessive autonomy as the root causes of excessive agency, and advises limiting the tools agents may call "to only the minimum necessary".¹
- The MCP specification says tools "represent arbitrary code execution and must be treated with appropriate caution", and that MCP itself "cannot enforce these security principles at the protocol level".²
- OWASP's Top 10 for Agentic Applications for 2026 lists tool misuse (ASI02), where agents bend legitimate tools into destructive outputs.³
- For broker-dealers, FINRA's 2026 oversight report added a new GenAI topic that covers AI agents, naming risks that include agents acting beyond the user's actual or intended scope and authority.⁴
Picture an agent with one job: prepare a mortgage file. It can also send a payment and open the HR files. Nobody planned that. It is simply what the agent was connected to.
Each agent gets only the tools it was given
Put Argent between your agents and your MCP servers. In Enforce mode, an agent can use a tool only if it was given that tool.
- Refused by default. A tool call is refused unless the agent holds a permission naming that server and tool, or the whole server. A permission never reaches across tenants.
- Listed equals callable. Argent shows each agent only the tools it can use. If the list cannot be worked out, the agent gets an empty list, never the full catalog.
- Unregistered servers are refused. A tool server nobody registered cannot be reached through Argent.
- Agents never hold the server's secret. When a server needs a credential, Argent fetches it from the vault at call time; if it cannot, the call is refused before anything is sent. See Keyless agents.
- Every agent that works through Argent. Argent governs the MCP traffic routed through it, and lists MCP servers your agents use that nobody registered, so each can be approved or denied. See Shadow AI discovery.
How it works, step by step
- Register. An admin registers each MCP server, or pulls servers in from a registry Argent connects to: the Official MCP Registry, the Docker MCP Catalog, or any registry that speaks the same format. Imported servers arrive switched off and pending approval.
- Measure. Argent reads each server's tool list and records a hash of every tool's definition. A tool that is new, or whose definition changed after approval, waits for a decision.
- Grant. Each agent gets tool permissions directly, or through a role, a group or an identity-provider mapping.
- Ask. The agent connects to Argent's one MCP endpoint and asks for its tools. It gets only the ones it can call, each named
server::tool. - Check. On every call Argent checks, in order: is the server registered and on, is the tool approved, has an admin switched it off or blocked this agent, and does the agent hold the permission. Dangerous-looking arguments, such as command injection or path tricks, are refused too.
- Answer. An allowed call goes to the server that owns the tool. A refused call returns a clear refusal to the agent, with a policy id and a trace id but never the rule text, and nothing reaches the tool. (In Watch mode the call goes through and the would-refuse is recorded.)
- Record. Every call, allowed or refused, lands on the signed record with the server, the tool, the reason and a hash of the arguments, never their raw values.
What your team sees
Day-to-day control lives on one console page, MCP Gate: "Every MCP server and tool your agents can reach."
| Where | What it shows |
|---|---|
| Catalog | Each server and its tools with a state: ON, OFF or NEEDS DECISION, how each is on (EVERYONE or BY POLICY), and a summary such as "2 servers, 6 tools: 4 on, 1 need a decision, 1 off". |
| Access editor | DEFAULT FOR EVERYONE: OFF, BY POLICY ("nobody gets it automatically; only the people and agents you name below can use it") or EVERYONE. Then EXCEPTIONS: WHO ELSE CAN USE IT, with ALLOW, NO RULE or DENY for roles, groups, agents and people. |
| Needs a decision | On a server in enforce pin mode: "New tools and tools whose definition changed stay blocked until an admin approves or denies them." APPROVE or DENY each one. |
| Who can call it | For each tool, who holds it by name, and how: a direct grant, a role, a group or an identity-provider mapping. |
| Activity | Per tool, calls and denied calls in the last 24 hours, and the three most recent callers by name. |
| Refusal reasons | In plain words: "No matching MCP scope", "Denied by a rule", "MCP server not registered". |
Admins decide, operators can ask
You choose who may use each server or tool: nobody, everyone, or only those you name. A block always wins.
- Who can change access. SuperAdmins and TenantAdmins change access on the MCP Gate page. Operators can deny and request; org and department admins act inside their own unit; auditors read. The page tells a lower role why a control is unavailable.
- Ask for access. REQUEST A SERVER, or request access to a tool, and an approver decides through the same controls.
- A block always wins. In the editor's words: "Allow lets one in; deny blocks one, and a block always wins over the default." A tool switched OFF cannot be used by anyone, whatever the exceptions say.
- Publish what you approved. Argent can serve its own approved servers as a registry, behind tokens an admin mints and revokes, and generates copy-paste client settings for Claude Code and VS Code that point them at Argent.
- Watch first. In Watch mode, calls Argent would refuse go through unchanged and the would-refuse is signed on the record, so you can tune before you enforce. See Watch before you enforce.
Where it fits in your program
Tool permissions are one layer of the same checkpoint that governs every request routed through Argent.
- Goals: Agent identity and ownership, Threat defense, See all AI in your company, Data loss prevention, Audit and accountability, Compliance evidence. All goals
- Use cases: Controlling agent tool access: MCP and agent-to-agent, Securing AI coding agents, Governing customer-facing AI assistants, Rolling out copilots and chat safely, AI in regulated and air-gapped environments.
- Industries: Financial services, Insurance, Healthcare, Public sector, Retail. All industries
- Works with: Identity tethering, Owner offboarding, Agent behavior watch, Prompt injection defense, Session replay, One-click audit evidence. All capabilities
How a team starts
Watch first. Nothing changes for staff until you decide.
- Register. Add your MCP servers, or pull them in from a registry; each imported server arrives off until an admin approves it.
- Switch the MCP Gate to Watch. Argent then records every call it would refuse while calls pass unchanged.
- Point clients at Argent. Use the generated client settings so agents reach their tools through Argent's MCP endpoint.
- Grant and enforce. Give each agent its tools, switch on refusals, and approve new tools as they appear.
- Inside your walls. Argent runs in your Kubernetes cluster, as a single-node bundle, or as an air-gapped bundle verified before import. Compare deployment options