Skip to content

Use casesAgents

Agents that you can trace, limit and stop.

An agent run is dozens of model calls and tool calls. Send them through one gateway key and one trace id: every step on one timeline, every tool from a definition you approved, and a spend limit the run cannot pass.

Start 14-day trialRequest a demo
Trace checkout-123 · triage agentExample
Step 1 · modelclaude-sonnet-5.5 · 200 · 1.8 s
Step 2 · toolgithub_list_issues · 200 · 0.6 s
Step 3 · toolgithub_close_issue · refused, not granted
Step 4 · modelclaude-sonnet-5.5 · 200 · 1.1 s

Example run. INFRO executes no tool; your MCP servers and their accounts stay yours.

The problem

Without a control plane

  • Each agent holds every tool server's token in its own config.
  • A tool's definition can change on the server and the agent follows it.
  • A run that loops is noticed on the invoice.

With INFRO

  • Tool credentials stay with INFRO; agents send their gateway key.
  • Tools are served from definitions an admin approved and pinned.
  • Each agent key has a limit, and every run is one trace.

How it works

  1. Register your tool servers once

    An admin adds each MCP server with its token and approves the tools. Agents never see the server's credential.

  2. Grant per key

    Denied by default: give an agent's key one server's read tools, all of them, or a named few. Rotating the key keeps its grants.

  3. Send one trace id

    Model calls and tool calls carrying the same X-INFRO-Trace-Id come back as one run from GET /v1/traces, with outcome, latency and cost per step.

What you get

Each one is documented, and each says which plan includes it. Where no plan is named, it is on every plan.

  • One trace per run

    Model and tool calls with the same trace id are returned as one run, each step with its outcome and latency.

    Docs

  • Approved, pinned toolsFrom Business

    Agents get the definition an admin approved; a tool that changes on the server is refused until someone reviews it.

    Docs

  • Grants per keyFrom Business

    Denied by default. Grant a key one server's read tools, all of its tools, or a named few.

    Docs

  • A limit per agent

    The agent's key has a ceiling the run cannot pass; the next request is refused before the provider call.

    Docs

  • GuardrailsFrom Business

    Pattern checks on prompts, responses, tool arguments and results: block, redact or flag. Detection can miss things.

    Docs

  • Kill switches on every plan

    Disable a server, block a tool or stop all tool traffic in one audited action. Turning access off is never plan-gated.

    Docs

Who it is for

  • Agent developersOne timeline per run instead of five dashboards, and tool access defined outside the agent's code.
  • Platform teamsTool credentials in one place, granted per key, with a log of every call.
  • SecurityA tool definition cannot change under an agent without review, and every call can be switched off.

FAQ

Does INFRO run my tools?

No. INFRO forwards each call to your own MCP server with the credential you registered, and records it. The servers and their accounts stay yours.

Do guardrails stop prompt injection?

No. Pinned definitions, grants and guardrails reduce what an injected instruction can reach; they do not detect or prevent injection itself.

Which agent frameworks work?

Any MCP client that speaks Streamable HTTP and sends an Authorization header, including the OpenAI Agents SDK and the Claude Agent SDK. Model calls use the OpenAI or Anthropic shapes they already send.

The full contract is in the MCP gateway docs, and every other question is on the FAQ.

Field guides

Related

All use cases

Get started

Try it on your own provider accounts.

Connect a provider, create a gateway key and change one base URL. Your providers keep billing you directly; 0% markup on model spend.

Start 14-day trialRequest a demo