does future agi block tool calls

How Does Future AGI Block Tool Calls?

Future AGI's guardrails for agent actions are part of the Agent Command Center (ACC), its OpenAI-compatible gateway to more than 100 model providers. The docs say gateway guardrails "run on every request and response flowing through Agent Command Center", catching unwanted content "before it reaches the LLM (pre-processing) or before it reaches your users (post-processing)."

• Tool Permissions validates tool and function call permissions at the pre-processing stage. Future AGI's May 2026 post puts the idea in one line: it "enforces which tools an agent may invoke, at the moment it tries", so that "a summarization agent should be structurally unable to call a payments tool."
• MCP Security validates MCP traffic at the same stage. The post says it "inspects that protocol traffic at the gateway: the calls the agent makes to MCP servers and the responses it gets back."
• Enforcement modes. Enforce blocks the request with a 403; Monitor lets it through with a warning logged; Log records the violation silently.

The same post is clear about location: "Agent runtime guardrails run at the gateway, the single network hop between your app and the model providers." To use them, you change your client's base_url to the gateway and swap the API key.

Which Tool Calls Can the Gateway See?

A gateway is a proxy on the path between your application and the model. It sees what crosses that path: the prompt, the tools offered to the model, the tool calls the model proposes in its response, and the tool results sent back on the next request. That is a good place for policy, because one configuration covers every agent that uses the gateway, in any language, with no code change beyond the URL.

It also means coverage is a routing question. The guardrail governs the traffic that goes through ACC. Before relying on it, check four things about your own setup:

1. Does every agent's model traffic go through the gateway, or do some services still hold direct provider keys?
2. Are your agent's tools declared as function calls that cross the gateway, or does a framework run some actions on its own?
3. Do your MCP calls pass through ACC, so MCP Security sees them?
4. For coding agents such as Claude Code or Codex: can that harness be pointed at the gateway, and does a check then apply to its shell commands and file edits? Future AGI's docs do not say, so ask them with your harness named.

What Happens to Actions Outside the Gateway?

A coding agent may execute a shell command or edit a file locally after the model proposes the action. If that action does not pass through Agent Command Center, a gateway rule cannot inspect it. Check this with the exact agent harness and tool path you use.

Failproof AI works at the agent hook layer. It uses code-based and LLM-based evaluations to find failure modes across production sessions, then lets teams turn a finding into a tested policy that can steer or block the action before the harness executes it. This complements gateway protection when agents also take local actions.

What Are the Options for Blocking Tool Calls?

• Failproof AI - for automatically finding failure modes and turning them into tested policies that steer or block actions inside supported agent harnesses.
• Future AGI Tool Permissions and MCP Security - for agents whose model and MCP traffic you route through ACC. One configuration, every routed request.
• An allowlist in your own tool dispatcher - if you wrote the agent loop, the code that executes a tool call can refuse names or arguments it does not expect. Free and exact, and you maintain it. Future AGI's protect() SDK call can screen an input from inside that code too.
• Galileo Agent Control - an Apache-2.0 project with pre- and post-execution checks on LLM and tool inputs and outputs, and deny, steer or log actions. You deploy and run it.

These are not exclusive. A company with a customer-facing agent behind the gateway and engineers running Claude Code has two request paths, and each needs a control on its own path. See Future AGI Protect vs Failproof AI policies for the side-by-side.

When Is Future AGI Enough on Its Own?

If your agents are applications you wrote, every model and MCP call already goes through the Agent Command Center, and the tools that matter are declared function calls, Future AGI's Tool Permissions guardrail in Enforce mode is the control you are looking for. Turn it on in Monitor mode first, read what it would have blocked, then switch to Enforce. Adding a second layer only earns its keep where actions happen outside that path.