MCPmed

arXiv:2507.08055v1 — "MCPmed: A Call for MCP-Enabled Bioinformatics Web Services for LLM-Driven Discovery" (Flotho et al., Saarland University / Helmholtz HIPS, Jul 2025). A position/initiative paper proposing that bioinformatics web services adopt the Model Context Protocol so agents can discover, invoke, and "verify" them — backed by a GEOmcp prototype, a cookie-cutter template, and an opensource hub (github.com/MCPmed, mcpmed.org). Call-to-action stage, not a finished system. Verified 2026-06-27 against the arXiv HTML.

Its framing

MCPmed argues that bioinformatics databases — GEO, STRING DB, the UCSC Cell Browser, PLSDB — are built for human browsers and fragile scraping, and should instead expose a "standardized semantic contract layered over existing API specifications" so LLM agents can use them directly. In their progression, a service moves up three layers — UI → API → MCP — where the MCP layer "explicitly associates each API endpoint with scientific concepts, along with versioned metadata, facilitating automated discovery, invocation, and verification of webservices." They ship a prototype (GEOmcp, exposing search_geo_series, search_geo_profiles, search_geo_datasets), a cookie-cutter template for new MCP PyPI packages, and a call for "a rigorously curated MCP app store" of "secure, pre-vetted MCP packages." It is explicitly "A Call for" adoption: an enabling-infrastructure vision with seed implementations, not a built platform.

Strengths relative to our approach

The enabling-infra vision is real and well-aimed, and the Foundry benefits from exactly this layer existing:

• A concrete, standards-based push to make services agent-actionable. Rather than each agent re-deriving how to call GEO, MCPmed wraps the endpoint once with a typed contract — "standardized manifests surface details such as rate limits, authentication, and pagination clearly, eliminating ad-hoc code and fragile scraping routines." This is genuinely useful plumbing.
• A working prototype, not just a manifesto. GEOmcp demonstrates an agent contextually refining a query ("microglia depletion mouse" → "CSF1R inhibitor") and recovering real datasets — the wiring works.
• Community scaffolding. A cookie-cutter template, an opensource hub, and a proposed curated app store lower the cost for service maintainers to join.
• Call-side provenance. "Concept-level versioning and audit-ready parameter tracking" record which endpoint, which version, which parameters produced a result — useful reproducibility metadata at the service boundary.
• MCP is itself a portability standard — one wrapped service is reachable from any MCP-speaking client, not bound to a single agent.

Where it sits on our values

Value	MCPmed
Produces skills	No — it emits MCP server packages that expose web-service endpoints as agent-callable tools, not portable, self-contained skill artifacts. The deliverable is a live tool surface, not a SKILL.md.
Progressive disclosure	No — the semantic metadata and manifests exist for machine discovery and invocation; there is no layered surface authored for a human to read and learn from.
Traceability	Partial, at the call boundary — "explicit provenance," concept-level versioning, and audit-ready parameter tracking trace what service call was made with which version/parameters. That is provenance of the invocation, not lineage from an authoritative knowledge source to an artifact.
CLI instructions derived from CLI docs	N/A — MCPmed wraps web services / REST APIs, not command-line tools. (Its sibling BioinfoMCP is the CLI→MCP case.)
Portability	No skills produced — but worth a nuance: MCP is itself a portability standard, so a wrapped service is callable across MCP clients. There is no runtime-agnostic skill artifact, though; portability is of the tool endpoint, via the protocol.
Human scrutiny	Minimal — the human-facing layer lives at the underlying web service's own UI; the MCP output is agent-facing. Nothing here is authored for a human to read, correct, or contribute to as knowledge.
Knowledge-base backed	No — it exposes external databases as live, queryable tool endpoints; the knowledge stays remote and is fetched per-call. There is no inspectable knowledge base beneath the wrapper.

Where it diverges from the Foundry

MCPmed strengthens the wiring layer — making bioinformatics services agent-actionable through a typed, discoverable contract — which is genuinely valuable and is a layer the Foundry happily sits atop rather than competes with. The precision that matters is the word "verification." In MCPmed it means verification of the webservice and the call: the endpoint exists, advertises a versioned contract tagged with scientific concepts, and records audit-ready parameters so a run is reproducible — "verify results through explicit provenance" is provenance of what was invoked, not of whether the analysis is sound. That is contract/availability/provenance verification of a service call. The Foundry's referee gate verifies the orthogonal thing one layer down: whether the statistical method beneath the result is valid — assumptions met, no double-dipping, the named method real and appropriate, error rate actually controlled. MCPmed can faithfully record that a confounded or invented method was run, with perfect call-provenance, and never flag it; method validity is outside its scope by design. It is a vision paper for the tool-exposure layer LLMs already use well — and the Foundry's edge is the validity gate beneath the tool call, not the tool's exposure. Distinct, too, from knowledgebase-mcp: that cluster exposes knowledge retrieval (RAG) to agents; MCPmed exposes web-service tools. Crux: MCPmed verifies that the call was well-formed and well-recorded; the Foundry referees whether what the call computed is statistically true.

Sources

• https://arxiv.org/html/2507.08055v1 — the position paper: the MCP-wrapping proposal, the UI→API→MCP layering, the GEOmcp prototype and its three search tools, the cookie-cutter template / hub / app-store call, and every use of "verification"/"provenance" (grounds the contract-vs-method-validity distinction).