Arch MCP Gateway

What Arch MCP Gateway does

Arch is an AI-native gateway intended to sit between an application, its language models, specialized agents, and Model Context Protocol tools. Search evidence describes it as a high-performance Layer 7 proxy based on Envoy. Its practical job is routing: inspect a request, direct it to the appropriate agent or MCP tool, provide one path to model providers, and expose traffic for monitoring. Arch also advertises clarification of vague requests, which can be useful before an agent invokes a tool with side effects.

That architecture is different from an agent framework. Arch does not replace the reasoning loop, business logic, or tool implementation. It provides the shared network boundary those components pass through. A platform team could use it to prevent every application from separately implementing model selection, prompt-to-tool routing, retries, logging, and policy. Compare that role with LiteLLM and Portkey, then read the MCP builder guide for protocol context.

Pricing and deployment reality

The required curl fetches of archgw.com and archgw.com/pricing returned no usable HTML. DuckDuckGo results surfaced public GitHub projects and third-party descriptions, but those are not sufficient to verify a current hosted plan, commercial support price, allowance, or license. Pricing is therefore recorded as not publicly verified, and this profile is flagged for manual review. Do not infer “free” merely because a repository appears in search; verify the repository owner, license file, release activity, and enterprise support terms before adoption.

Where it fits

Arch makes sense when several agent applications need the same routing and governance. A concrete pilot is to connect two models, two MCP servers, and one specialized agent. Send 100 representative requests and measure correct-route rate, added latency, retry behavior, trace completeness, and what happens when a tool is unavailable. Confirm that authentication reaches the downstream service with the intended user identity rather than a broad shared credential.

The upside is architectural consistency: routing, telemetry, and policy live in one place. The downside is another production component that can become a bottleneck or single point of failure. Teams must operate upgrades, capacity, TLS, secrets, and fallback behavior. Prompt-based routing can misclassify unusual requests, so consequential tool calls still need deterministic authorization and approval.

Arch is worth investigating for platform engineers standardizing many agents, but inaccessible vendor pages prevent a confident purchasing recommendation today. Compare its MCP role with Anthropic MCP and Smithery, and apply AI agent security best practices before exposing write-capable tools.

Final evaluation check

Before committing, assign an owner, a four-week test period, a baseline, and a measurable pass threshold. Record successful completions, failures, retries, human correction time, total fees, and downstream usage costs. Recheck the vendor price and contract on the purchase date because features, allowances, and promotional terms can change after this research date.