Paper2Agent converts a published paper's code into an MCP server a chat agent can query

Paper2Agent writes its own tests. The pipeline analyses the paper and its codebase with multiple agents, builds the model context protocol server, then generates and runs tests to refine that server and increase its robustness ^[1]. A test written by the same pipeline that wrote the wrapper can show that a tool executes and returns what the paper returned. That is a narrower check than whether the tool suits a dataset the paper never saw.

The authors locate the failure of reuse in setup. A reader has to find the code repository, install dependencies, configure environments and work out the correct inputs and outputs ^[8]. "Even with well-maintained repositories, this process is often non-trivial," they wrote ^[9]. For AlphaGenome the friction is itemised: install the environment, create client objects with API keys, construct inputs such as variant objects, select the output modalities, and understand the API hierarchy and parameter semantics along the way ^[10]. An agent removes that setup. Which output modality answers the biologist's question is still a judgement call.

One case study goes past reuse. Several Paper2Agent agents worked together to prioritize a causal gene for psoriasis ^[6]. The answer a user gets depends on the model doing the calling as much as on the code being called: the route to that output runs through a chat agent, because the paper MCPs are connected to a client such as Claude Code, which invokes the paper's tools in response to natural language ^[2].

Three tools appear in the case studies ^[11], against a claim pitched at the whole literature: the authors write that any research paper can be converted into an agent that embodies the knowledge and methods described in the publication ^[14]. The validation reported in the abstract is that the agents reproduce the results of the original papers and carry out novel user queries ^[3]. The abstract does not state how many papers the pipeline was run on, or how often a conversion failed ^[12]. Those two numbers would say whether the framework also works on a paper whose code has not been run since acceptance.