Introducing the Papers with Code MCP server! To celebrate its release, I've asked Claude Code to research the architecture of Jev by TypeSafe AI using the `search_papers` tool. Here's what it came up with: "If Jev is architecturally anything public, it's a large schema-conditioned bidirectional encoder with parallel label-query... heads (GLiNER/ML-Decoder shape) trained with a proper-scoring-rule RL objective (RLCR shape), scaled far past the ~150M-param range those papers operate in — "neither small nor an LLM" fits that. The 40–200× speedup is consistent with removing autoregression, not with any exotic mechanism."show more

Find it here: In building this, I realized the same thing as @trq212 and @RhysSullivan already noted: an MCP might be easier to work with than a CLI. This now also powers the chat interface on Papers with Code!

@typesafeai Super cool, you should publish this on the @swarms_corp marketplace so builders can access it easily

@typesafeai I'd love the output to separate 'the paper says this' from 'the agent thinks these architectures are related'. Both are useful, but that distinction would make a research summary much easier to trust and check.

@typesafeai The bigger implication is that MCP can turn paper discovery into an interactive research pipeline: find sources, extract claims, compare architectures, and preserve citations in one agent workflow.

@typesafeai Ouch!

@typesafeai This is a clean tool boundary: MCP finds relevant papers, then Claude forms the architecture hypothesis. A claim-to-source map would make the result even stronger, especially for the speed explanation, because it separates supported evidence from plausible inference.

@typesafeai Searching papers by hand is the slowest part of agent work. Does it index the full archive or just recent ones?

@typesafeai Papers with Code went from a curated list to an MCP server. The library card became the API endpoint. we used browse. now we query. same books, faster shelf talkers

@typesafeai The useful shift is that the model can now interrogate a corpus instead of just autocomplete around it. The next bottleneck is provenance: every retrieved claim needs a stable paper, version, and citation trail or the speed just scales confident errors.

@typesafeai Does the server expose paper metadata and code links as structured results, or leave Claude to parse them from text?

@typesafeai Yes, it also exposes the metadata of each paper (such as authors, organizations, GitHub/project page links, tagged tasks and methods, evals, etc.)

@typesafeai An MCP that pulls Papers with Code into the agent loop is distribution for research, not another bookmark. Claude Code researching architecture on command is the demo. The durable win is citations and code that stay attached to the claim.

@typesafeai A search tool over papers makes the agent argue from sources instead of from memory, which is where most confident architecture claims go wrong. Does it handle papers that have no code attached?

A Papers with Code MCP server is a great fit for research agents because the source trail can stay attached to the answer. I’d make citations and search queries first-class outputs, so a surprising architecture summary is easy to verify instead of becoming a polished hallucination.