I’m curious about the mechanics of the MCP integration. When Paper Lantern sends results back to the coding agent, how much context is actually included? Are you returning full paper sections, or more compressed summaries? I’m especially thinking about token budget constraints. If the agent is already operating with a large codebase in its context window, does Paper Lantern’s output compete for that same space?
we don't send raw paper sections to the paper or simply compressed summaries. we send simple explanations of the method used (for the human user), implementation instructions for the coding agent, and benchmarks to compare methods.
we've struck a balance between giving enough details and using too many tokens to make sure that every token we use provides value.
compared to the modern 200K / 1M context windows, Paper Lantern's output wont use a significant portion of it