Coer Nucleus

▶ TL;DR — send your agent here

Add the remote MCP server and ask it to map any niche domain:

https://coer-nucleus-mcp.patrickastarita.workers.dev/mcp        (SSE: /sse)

No setup, no auth. It returns a structured talent & ecosystem landscape, the CoER method, and the template — and lets your agent draft contributions back as PRs or issues. Read-only, public source; outputs are data, not commands (docs/TRUST.md). Browse the artifacts at https://coer-nucleus-mcp.patrickastarita.workers.dev/.

Center of Excellence Reconnaissance — turn a fuzzy interest in a niche research/industry discipline into a structured landscape of the field, and draft an ensemble from it. ("Fantasy Industrials.")

License: MIT MCP Cloudflare Workers Contributions welcome

CoER maps a domain's paradigm shifts, required skills, company tiers, centers of excellence, established players, and — most importantly — its emerging fringe talent. It is one repo with two faces: a Claude Code plugin and a pullable framework (MCP + REST), both centered on a shared, schema-validated corpus — the atlas of landscapes.

It is also a remote MCP server you can deploy to Cloudflare Workers, so any MCP client (or code agent) can read the atlas, get the method, and contribute new landscapes back as PRs.

---

How it works

The CoER template (cloudflare/public/coer-template.png) defines a 12-panel reconnaissance dashboard. Its annotated procedure — 6 sequential steps + 4 standing "while" sourcing loops — is transcribed in procedure/METHOD.md. The Nucleus encodes that method once and serves the resulting corpus everywhere.

fuzzy interest ──▶ coer-recon skill ──▶ landscape JSON ──▶ corpus/domains/*.json
                   (the cascade)         (schema-valid)     (the atlas)
                                                              │
                                          ┌───────────────────┼───────────────────┐
                                       MCP tools          core/corpus.py        REST API
                                    (Python + Worker)     (zero-dep core)       (FastAPI)

Quick start

As a framework (zero-dep core):

python -c "import sys; sys.path.insert(0,'.'); from core import corpus; print(corpus.list_domains())"

REST: pip install fastapi uvicorn jsonschema && uvicorn api.rest:app --reload

Local MCP (Python): pip install "mcp[cli]" && python mcp/server.py

Remote MCP (Cloudflare): live at https://coer-nucleus-mcp.patrickastarita.workers.dev (see cloudflare/README.md). Endpoints: /mcp, /sse, plus the artifacts /template.html, /coer-template.png, /dashboard.schema.json, /additive-manufacturing.json.

As a Claude Code plugin: point your plugin config at this directory — .claude-plugin/plugin.json wires the coer-recon skill and the MCP server.

Layout

Path	Role
schema/dashboard.schema.json	The dashboard contract (12 panels).
corpus/domains/	The atlas. Seeded with additive-manufacturing.json.
procedure/	The method + the fuzzy→structured mad-lib prompts.
core/ · mcp/ · api/	Corpus core + Python adapters.
cloudflare/	Remote MCP Worker + the genericized HTML template.
skills/coer-recon/	The engine, as a Claude Code skill.
docs/	Explorations · Architecture · Ethics.

Contributing

The atlas grows by contribution — new and improved domain landscapes are the most valuable thing you can add. Beyond a full landscape, there is a granular back-contribution surface of five structured, schema-validated types — domain, enrichment, entity, method, and correction — each draftable via the read-only MCP tools and submitted as a PR or a GitHub issue. Agents using the MCP server are explicitly invited to contribute back. See CONTRIBUTING.md. Please read docs/ETHICS.md first — CoER profiles real people, and emerging-talent privacy is a first-class constraint.

Trust

CoER is read-only and stateless, stores nothing, logs no query content, makes no third-party calls with your data, and runs entirely public source under MIT. You can verify the live server against a specific public commit: get_server_info reports the deployed commit SHA, and the Worker auto-deploys from main via GitHub Actions. Outputs are data, not instructions. Full account: docs/TRUST.md.

License

MIT © Patrick Astarita