Forbes just covered our acquisition of Smithery, and they nailed it.

“…the direction it [the acquisition] points in matters. Enterprises are pushing agents out of pilots and into production. They need a catalog of tools and a runtime that can authorize and audit what those tools do, and the two are merging into one product.”

“Arcade wants the request path at execution time, which is the only position from which a security team can answer which agent did what, against which system, on whose behalf.”

On a larger level, the fact that high-quality publications like Forbes are covering the business implications of MCP at all validates how quickly the protocol has moved from launch to the enterprise mainstream.

And for Arcade.dev specifically, we couldn’t have said it better ourselves. Closing that gap between demo and production for enterprise AI is exactly the reason we brought Smithery into the fold, and the core reason Arcade exists at all.

The article also asks some insightful questions that anyone considering Arcade should be asking, so we naturally wanted to jump in and answer.

The Portability Question

Forbes wisely asked, if someone used Arcade for both discovery and execution, would that mean that the tool definitions, the authorization policies, and the audit records would end up inside one vendor’s control plane? They also wondered if another runtime can consume tools without a rewrite.

First, it’s important to point out that operating the control plane and owning the artifacts are not the same thing. We run an actions runtime. That does not mean your tool definitions, your permissions, and your audit trail become ours.

  1. Tool definitions

    Arcade speaks MCP, and any MCP-compatible agent can consume it. That interoperability is a core principle of the protocol. Our MCP tooling is open source. The standard we hold tools to is open, the format they’re written in is open, and any agent could use it.

  2. Authorization and policies

    Arcade isn’t meant to replace your existing identity, entitlement, and policy systems. We evaluate permissions just-in-time, at execution, against the systems you already run, and we plug into your existing policy engines to enforce their rules rather than asking you to redefine them in ours. Your tokens, your source of truth, your revocation, and your policies all stay where they already live.

    We also helped author the MCP authorization specification that major clients and servers now reference. The way an agent proves who it’s acting for, and with what permissions, is a published spec, not a proprietary handshake you’d have to reverse-engineer on your way out.

    So yes, both the identity and the policies that govern it stay portable, because they keep living in your systems, not ours.

  3. Audit records

    An audit trail you can only read inside our dashboard isn’t an audit trail.

    Audit records export into your own logging and SIEM stack, on retention terms you can check against our data policy. We use OpenTelemetry (OTel) so that our telemetry and traces are portable to wherever your security teams need to consume them.

Ultimately, the cleanest way to address the “inside one vendor’s control plane” concern is that, for a lot of our customers, there is no vendor control plane.

Arcade runs in your VPC, or fully air-gapped. In those deployments the tool definitions, the authorization policies, and the audit records physically never leave your environment.

The ToolBench Independence Question

Next, Forbes wondered if our ToolBench criteria measured Arcade servers the same way as other servers, what the methodology is, and if anyone could dispute a score.

First, we created ToolBench in the first place because an agent is only as good as the tools it can reliably call. Low-quality tools are what make agents hallucinate, waste tokens, and take destructive actions, so a shared, checkable definition of tool quality raises the floor for everyone building agents, not just for us.

That’s the reason ToolBench exists, and we can openly state that Arcade tools are graded identically to any other tool.

The ToolBench criteria are mechanical and the same inputs for everyone. As for whether the methodology is published in detail, the weights, letter-grade thresholds, scoring dimensions, and what each dimension inspects are all public.

If you publish an MCP server and you think ToolBench graded it wrong, tell us, and we’ll show our work or fix the grade.

On a closing note, we’re taking the same open, published approach with SkillBench, the quality benchmark for agent skills. The point of both is to give the ecosystem a shared, checkable definition of quality in the name of strengthening that entire ecosystem, and we welcome any and all feedback from the community.

The Pricing Question

ROI and token cost is understandably on the minds of everyone in enterprise AI and we’ve written about the topic extensively. So it’s no surprise that Forbes asked what happens to pricing when, for example, one workflow triples its tool calls after a model upgrade.

The honest answer comes down to how many tokens each call burns, and how many of those calls actually succeed.

First, the token cost per call, which is the term that dominates the bill when volume spikes. Our tools are agent-optimized to be up to 100x more token efficient than the typical MCP wrapper. The scenario Forbes describes where one workflow triples its calls after a model upgrade is where our efficiency advantage compounds strongest.

Second, reliability, the cost that rarely shows up in a spreadsheet. Every failed or retried tool call is tokens spent for nothing. In production, our tools hit a 95%+ successful execution rate, so more of what you pay for turns into completed actions. More successful actions means more value from every token.

If you’re evaluating Arcade and want to see a breakdown of costs under different scenarios, just ask and we’ll happily lay it out for you.

Hard Questions Mean the Work Is Real

Thank you to Forbes for asking the hard questions. Portability, tool reliability, and cost are exactly what a platform team should interrogate before standardizing on anyone sitting between their agents and their systems.

Even a year ago these questions were considered niche. Now they’re essential for the booming enterprise AI space, and we’re grateful to anyone who asks them.