Quarkus Insights #259: MCP Servers Without the Boilerplate — Introducing reShapr

As the MCP ecosystem matures, a new pain point has emerged: teams are building a sprawling, inconsistent collection of hand-crafted MCP servers — one per use case, each with its own security model, its own stack, and its own operational burden. Laurent Broudoux, co-founder of the Microcks CNCF incubating project and member of the Open Innovation and AI team at Postman, joins hosts Eric Deandrea and Holly Cummins in episode 259 to introduce reShapr: a declarative, platform-level approach to turning existing APIs into production-grade MCP servers — without rewriting them.

Eric opened with a brief news round: work on Quarkus 4 is actively progressing. The working group board is the best place to follow along. A candidate release for Quarkus 3.40 is expected this week, which is intended to be the last LTS on the 3.x stream. Eric will also be at Developers North America in San Jose, California the following week, demoing testing of non-deterministic things with Langfuse and the Quarkus integration.

Laurent framed the talk with a recurring observation from large enterprise Microcks adopters: MCP servers are proliferating rapidly, but not in a healthy way. Every new use case spawns a new hand-built MCP server, each one consuming engineering time, requiring deep knowledge of a fast-moving protocol (major versions every six months), and inevitably implementing security in a subtly different way from the one built the sprint before.

The protocol itself is broad and does not enforce consistent usage, so even experienced engineers can produce servers that flood an LLM with far too many tools and far too much context — the exact condition that causes hallucinations and failed orchestration. Laurent showed a concrete example: importing the full GitHub GraphQL schema into an MCP server exposes 283 operations. In practice, an LLM cannot reliably pick the right one.

reShapr is built around the observation that most MCP servers are just translations of existing APIs. Rather than generating code for each translation, reShapr executes translations on the fly from a declarative description. The workflow is:

No compilation step, no deployment pipeline. The configuration plan is the server.

Laurent walked through the reShapr dashboard live. Starting from a remote Open-Meteo OpenAPI URL, he imported the spec, skipped authentication for simplicity, created a configuration plan, and had a running MCP server in under two minutes. Connecting the server URL to Postman’s built-in MCP inspector confirmed that the full OpenAPI description had been translated into a browsable, callable MCP server — including a working forecast endpoint returning real results.

Exposing an API directly is the trivial case. reShapr’s core value is reshaping: fine-tuning what the MCP server exposes so that the context passed to the LLM is exactly what the workflow needs, no more.

For the GitHub GraphQL example, Laurent defined a custom MCP server that surfaces only two operations: get user with latest followers and compare two users. Both are higher-level than anything in the raw schema — they are domain-shaped tools for a specific agentic workflow.

Reshaping capabilities include:

Looking ahead, the DSL for orchestration is designed to be pluggable: Apache Camel routes and Arazzo specifications are both on the roadmap as declarative alternatives to the current JavaScript snippets.

reShapr has first-class versioning support: multiple API versions can coexist, configuration plans can be versioned independently, and changes pushed through the UI or CLI are streamed to proxies in real time. Rolling deployments are supported via gateway groups, so updates can be staged rather than applied globally.

Critically, MCP protocol versioning is also handled automatically. The July 2025 version of the MCP spec introduced a sessionless model and a new multi-trip request pattern. reShapr negotiates protocol versions with clients, so when a new protocol version ships and reShapr adds support, all servers built on it upgrade automatically.

The security layer was demoed briefly. Each configuration plan includes a step for configuring MCP endpoint authentication — API key, OAuth (with support for SPIFFE, PKCE, and the new enterprise-managed authorization extension from recent MCP protocol versions), or no authentication for local development. The backend authentication to the underlying API is configured independently, allowing any combination of client-facing and backend auth schemes.

Some of this may sound similar to what API gateways do. Laurent’s positioning relative to AI gateways: reShapr operates at the application protocol level (translation and policy), not the infrastructure level (routing, rate limiting, LLM proxy). In practice it sits behind an AI gateway that handles discovery and LLM routing, with reShapr performing the protocol adaptation before touching the business API.

Laurent closed with a deep dive into the stack, opening with "reShapr is, I would say, mainly – totally built on Quarkus."

The architecture has two major components:

OpenTelemetry is woven throughout the proxy: per-call traces show the protocol negotiation overhead (measured in single-digit milliseconds), the parallel backend calls from orchestrated tools, and full audit logs when enabled on a configuration plan.

Benchmarks from the upcoming 1.0 release baseline show a 2.5 ms overhead per request at a nominal load of 150 req/s, and throughput approaching 4,000 req/s on small payloads — numbers Laurent noted that he had not optimized much yet. "That’s the Quarkus dream, isn’t it? You don’t even need to optimize your performance because you just get it in the platform," as Holly put it.

The business-perceived performance benchmark was the headline result: using the official GitHub MCP server versus reShapr-shaped custom tools for a quality-metrics workflow that fetches project info, PRs, reviews, and CI results, reShapr delivered 97% fewer tokens consumed, with dramatically fewer round trips and lower latency. For teams watching LLM usage costs, that single number is the talk’s most shareable line.

reShapr tackles one of the most pressing friction points in the current AI tooling landscape: the gap between the APIs that enterprises already have and the precisely-scoped, agent-friendly tools that LLMs actually need. By making the translation declarative and the execution platform-managed, it removes the expertise burden — MCP protocol depth, proxy engineering, security implementation — from every individual team and centralises it once. The Quarkus foundation means the runtime is fast and the operational profile is lean; the 97% token reduction benchmark suggests the architecture-level bet is sound. Next up — at least according to the closing exchange — a Quarkus extension for reShapr itself.