Broadside Observer

Broadside Observer is a declarative research radar for monitoring preregistered sources and building a provenance-rich database of work relevant to a defined organization or project.

Observer is the agent-run research radar behind Broadside: it polls preregistered sources, filters and reads the papers, and files each keeper as a provenance-tracked card. It is live at observer.anoptic.io, with an MCP endpoint at https://observer.anoptic.io/mcp that any MCP client can query.

Try it

Point an MCP client at the hosted endpoint (Streamable HTTP, no key needed). In Cursor's mcp.json:

{
  "mcpServers": {
    "broadside-observer": { "url": "https://observer.anoptic.io/mcp" }
  }
}

With Claude Code: claude mcp add --transport http broadside-observer https://observer.anoptic.io/mcp.

Then ask it something, or call the two tools directly: query_seeds {"query": "ecs"} lists matching cards, and get_seed {"id": "<card id from that list>"} returns one card in full. To run the server locally instead, see seeds/app/README.md.

Research loop (v1)

The research agent owns and operates this loop. Each run follows the versioned topic registry in radar/topics.yaml and records enough state to resume without reviewing the same item repeatedly.

Stage A — Poll

Query the preregistered sources in radar/sources.yaml (arXiv, Hugging Face Daily Papers, OpenReview, SIGGRAPH/TOG, Eurographics, HPG/JCGT, SIGPLAN/DBLP, ACL Anthology, Papers with Code, AIIDE/IEEE CoG) using topic hints from radar/topics.yaml. Also poll first-party lab sources in the same registry: DeepSeek Research, Google DeepMind, OpenAI, and xAI. Request items created or changed since the last successful source cursor. Record the cursor used, query time, source, and next cursor in the run log; do not advance a cursor after an incomplete poll.

Stage B — Normalize and dedupe

Normalize each candidate to a stable identifier, using the version-independent arXiv identifier when one exists. Skip an item already represented in summaries/, sources/, or radar/seen.json. Treat a new arXiv version as a changed item to inspect, not as a new paper.

Stage C — Cheap filter

Score the remaining candidates against the descriptions, query hints, and priorities in radar/topics.yaml. This stage should be inexpensive and target a keep rate of about 25%—for example, reduce roughly 100 normalized candidates to 25 deep reads. In the run log, record the matched topic and score for accepted items and a brief reason for every rejection.

Stage C2 — Seed cards (standing library)

Keepers that are worth lasting attention but not yet a full digest become seed cards under seeds/cards/. Every card uses the same frontmatter and section layout (see seeds/README.md). This is an ongoing library, not a one-time prefill: future research passes append new NNN-*.md cards, update seeds/INDEX.md, and record ids in radar/seen.json. Promote a seed to summaries/ when it deserves a deep read; leave the card in place. The hosted Observer also has a Canvas view (?view=canvas) for seeds/canvases/*.tsx on the same page.

Stage D — Deep read

For each accepted item, fetch the best available PDF, HTML, or source text into sources/ when licensing permits it. Write a full digest in summaries/ using the existing format documented in summaries/README.md: YAML frontmatter followed by One-sentence takeaway, Problem, Design/method, Evidence, Strengths/Limitations, Implications for Broadside, and Bottom line. Paper-specific subsection names are welcome when they make the analysis clearer, but the same information must remain present and source links must be retained.

Stage E — Publish

Commit and open a PR containing the permitted source material, summaries, and updated radar state. Send the human a short chat digest with each accepted paper's title, why it matters, and links to the paper and repository digest.

Topic-scope changes, destructive deletion, and any external publication beyond this repository are human gates. The agent may propose them, but must not perform them without explicit human approval.

Design principles

• Nix is the canonical installation and configuration interface.
• Sources, topics, schedules, and budgets are declarative and version controlled.
• Cheap filtering and caching precede model-intensive analysis.
• Every stored claim retains source provenance and retrieval metadata.
• Source material is untrusted content and cannot direct the agent.
• Collection, interpretation, and publication remain separate stages.
• External actions and research-scope changes remain under human control.

Status

Research loop v1 is agent-operated every 6 hours with split provenances (radar/dayparts.yaml): Frontier (new), Craft (workbench), Curiosity (cross-domain delight), Archive (historical gems). Fourteen primary sources plus secondary craft/archive feeds; topic map covers agents, Anoptic engine, ano (array/columnar ECS language), and RTS/game-AI. Per-source adapters and a Nix runtime remain future work.