A multi-step academic figure agent harness that enables AI agents to plan, generate, evaluate, and iterate publication-grade figures from sources like PMIDs, preprints, or freeform briefs.
Enables OpenAI-compatible clients to send llama-server only the MCP tools relevant to each request, reducing prompt size and cold prefill latency through pluggable selectors such as Laya or BM25 while preserving normal tool calling.