MCP server for autonomous overnight research loops.
search_literature: queries searchthearxiv for relevance-ranked paper discovery.search_literaturenow returnshealthmetadata with cache, retry, fallback, and empty-result signals.read_paper_text: fetches full arXiv HTML and converts it to plain text chunks.instruction://overnight-research-loop: strict loop protocol resource for coding agents.instruction://research-memory-template: markdown template for research memory files.instruction://research-evidence-template: markdown template for extracted evidence notes.instruction://research-session-checkpoint: compact live state buffer for compaction-safe continuation.instruction://research-task-queue: round-robin task queue for long-running sessions.instruction://research-stop-condition: bounded pause and resume guidance for exact stop targets.instruction://research-validation-rubric: ranking and validation checklist for evidence-backed ideas.overnight_research_bootstrap: reusable MCP prompt to start the loop.
The server is designed for restartable long sessions rather than literal infinite execution in one context window.
- Keep a small checkpoint in
instruction://research-session-checkpoint. - Keep a live round-robin queue in
instruction://research-task-queue. - Write one
research/memory file per idea. - Track query provenance and
healthmetadata so degraded discovery runs stay auditable. - Enforce search -> read -> implement by reading full paper text before coding.
- Refresh the checkpoint after meaningful progress and before large task switches.
- Treat compaction or timeout as a normal pause, then resume from the checkpoint and queue.
- Use
research/ideas/for per-idea notes,research/evidence/for supporting passages,research/checkpoints/current.mdfor the live state,research/queues/live.mdfor the work queue, andresearch/templates/for reusable forms.
npm run buildnpm run start
For development:
npm run dev