Arch Gateway (now Plano)

Arch Gateway is now Plano. The former official GitHub URL redirects to the katanemo/plano repository, whose current README describes Plano as an AI-native proxy server and data plane for agentic applications. This naming change matters: buyers searching for Arch Gateway should evaluate the maintained Plano code and docs rather than assuming an old Arch article represents the current product. The project is Apache-2.0 licensed and says it is built on Envoy. It moves recurring infrastructure—agent routing, model-provider handling, guardrail hooks, traces, and metrics—out of individual agent code and into a separately operated layer.

The integration model is more concrete than the broad AI gateway label. The official example defines agents, provider credentials, listeners, routing, and tracing in YAML. Each agent remains an HTTP service implementing an OpenAI-compatible chat-completions endpoint. Plano can route by model name, semantic alias, or preferences; its example uses a 4-billion-parameter Plano orchestrator to choose weather and flight agents. The README documents zero-code Agentic Signals plus OpenTelemetry traces and metrics, and filter chains for moderation, jailbreak protection, and memory hooks. Those are useful capabilities for teams that otherwise duplicate routing and telemetry in every service.

Pricing requires careful wording. Source is available under Apache 2.0, so there is no software license charge for self-hosting, but infrastructure, model inference, logging, and engineering remain real costs. The README says Plano-family models are hosted free in US-central for a first-run developer experience. Teams scaling to production are told either to run those models locally or contact the project for API keys. No numeric production API, support, or managed-service price was found on September 18, 2026. A production budget should not treat the quickstart allowance as an unlimited free tier.

For a pilot, reproduce the repository’s two-agent pattern with non-sensitive data. Measure routing accuracy over at least 100 labeled prompts, p50 and p95 added latency, completion rate, duplicate actions after retries, trace completeness, and model cost per completed task. Disconnect one agent, expire a provider credential, send a malformed streaming response, and verify timeout and fallback behavior. Inspect whether logs capture prompts or secrets. Set sampling intentionally instead of copying the example’s 100% value into production, and confirm that restarting the gateway does not lose state your workflow needs.

Plano is most credible for developer teams wanting an out-of-process routing and observability layer that they can operate. It is less attractive for a team needing a fully priced turnkey SaaS or unable to accept another network dependency. Before adoption, confirm release cadence, production support, hosted-model terms, security policy, and migration implications of the rename. The staged claim that Arch is both an MCP client and server could not be confirmed in the current README, so do not make it a purchase assumption without a working transport test and current documentation.