cre-skills-plugin

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

 ██████╗██████╗ ███████╗    ███████╗██╗  ██╗██╗██╗     ██╗     ███████╗
██╔════╝██╔══██╗██╔════╝    ██╔════╝██║ ██╔╝██║██║     ██║     ██╔════╝
██║     ██████╔╝█████╗      ███████╗█████╔╝ ██║██║     ██║     ███████╗
██║     ██╔══██╗██╔══╝      ╚════██║██╔═██╗ ██║██║     ██║     ╚════██║
╚██████╗██║  ██║███████╗    ███████║██║  ██╗██║███████╗███████╗███████║
 ╚═════╝╚═╝  ╚═╝╚══════╝    ╚══════╝╚═╝  ╚═╝╚═╝╚══════╝╚══════╝╚══════╝

        v5.2.1  ·  MCP reconnect patch — plugin-root entrypoint, Claude/Codex cache compatibility

                                │                                                                         |        
            *                   │           │           │                     _                          /|     |#|
           /|                  _│_         _|_          A          _        _[_]_                _|_    / |     |#|
          /_W_                /:::\       [_=_]        /_\        | |     _/_:::_\_      __     |:::|  |  |     |#|
          ( • )              /|:::|\      |:::|       ((^))       | |    |X/:::::\X|    |__|    |:::|  |  |     |#|
        \__/ \__[]           | |:::| |   _|_:_|_     /(:::)\     _|_|_   |/X:::::X\|    |::|    |:::|_ |  |     |#|
           | |              | |:::| |   |:::::::|    |:::::|    |:::::|  |\X:::::X/|    |::|    |:::::||  |    _|#|_
          /   \             | |:::| |  _|:::::::|_   |:::::|   _|:::::|_ |X/:::::\X|   _|::_|   |:::::||::|   |:::::|
         _|___|_            | |:::| | |:::::::::::|  |:::::|  |:::::::::||/X:::::X\|  |::::::| _|:::::||::|  _|:::::|_
    ___|=======|___________|_|:::|_|_|:::::::::::|__|:::::|__|:::::::::||\X:::::X/|__|::::::||:::::::||::|_|:::::::::|
           BROOKLYN BRIDGE                           WILLIAMSBURG BRIDGE                     QUEENSBORO BRIDGE
      _|_                     _|_                  |▔▔▔|               |▔▔▔|               _|_         _|_         _|_
     | ∩ |                   | ∩ |                 |XXX|               |XXX|              //|\\       //|\\       //|\\
    /| ∩ |\                 /| ∩ |\               /|   |\-------___---/|   |\            ///|\\\_____///|\\\_____///|\\
   / |___| \-------_-------/ |___| \-------------/ |___| \     /   \  /|___| \----------////|\\\\   ////|\\\\   ////|\\
  / /     \ \     / \     / /     \ \           / /     \ \   /     \/ /     \ \       / /  |  \ \ / /  |  \ \ / /  |  \
≈/≈/≈≈≈≈≈≈≈\≈\≈≈≈/≈≈≈\≈≈≈/≈/≈≈≈≈≈≈≈\≈\≈≈≈≈≈≈≈≈≈/≈/≈≈≈≈≈≈≈\≈\≈/≈≈≈≈≈≈≈\≈/≈≈≈≈≈≈\≈\≈≈≈≈≈/≈/≈≈≈|≈≈≈\≈/≈/≈≈≈|≈≈≈\≈/≈/≈≈|≈≈≈\
≈/≈/≈≈≈≈≈≈≈≈\≈\≈/≈≈≈≈≈\≈/≈/≈≈≈≈≈≈≈≈≈\≈\≈≈≈≈≈≈≈/≈/≈≈≈≈≈≈≈≈≈\≈\≈≈ EAST RIVER ≈≈≈\≈/≈≈≈≈/≈/≈≈≈≈|≈≈≈≈\≈/≈≈≈≈|≈≈≈≈\≈/≈≈≈|≈≈≈≈\

CRE Skills Plugin

A Claude plugin delivering a large library of commercial real estate skills covering the full investment lifecycle -- deal sourcing, screening, underwriting, structuring, due diligence, capital markets, market research, asset management, leasing, investor relations, development, disposition, tax planning, ESG, portfolio strategy, and daily property operations. Each skill includes structured process logic, reference documents, chain connections to other skills, and Python calculators for precise quantitative output. Deploys as a plugin in Claude Code (CLI or Desktop Code tab) or as a local MCP server for Claude Desktop Chat tab. See Key Stats for current counts and Release Maturity for status by surface.

How the skill taxonomy works for your team

Every skill now declares a classification that tells your team what it is and how much to trust its output. In plain terms: a micro or normal skill does one bounded job — normalize a T-12, size a loan, draft a quarterly LP letter — and is invoked as a single unit. An orchestrator is a conductor that runs several skills in sequence (an underwrite-to-IC pipeline, a document-to-database ingest); it is a workflow, not an atomic tool. A workspace is a business-facing surface that routes a whole domain — your fund team's LP reporting, your asset managers' operations cockpit — to the right specialist skill and holds context across sessions. For a deal team, an asset manager, or a fund team, the classification is not cosmetic: it determines which outputs are decision-grade and therefore need a human sign-off before they reach an investment committee, a lender, or an LP. Skills whose output can land on a board deck or a distribution notice are marked decision-grade and carry an explicit human gate (analyst review, AM/CFO sign-off, IC approval, or lender/investor review) plus a source posture that refuses to emit a final figure it cannot trace to your data — no fabricated model references, no unresolved $X placeholders in a final-marked report. Everything else is advisory screening output your operator validates. See docs/DATA_GRADES.md for the data-grade ladder and docs/integrations/amos-skill-manifest.md for how this metadata is exported to AMOS.

Open core

This repository is the free, open-source core of CRE Skills, licensed under Apache-2.0. It is the knowledge layer that runs locally in your Claude Code session: the skill library, routing, calculators, and the MCP surface. It stays free and open.

A separate, paid edition, cre-skills-pro, is in development as the institution-grade governance layer that composes on top of this core. It adds a deterministic control plane: lifecycle hooks, a four-eyes approval matrix, append-only redacted deal-state memory, an audit log, and a first-run setup TUI, so decision-grade output (IC memos, valuations, lender and LP deliverables) is gated, provenanced, and auditable for institutional use. It will be offered commercially; this free core is not going away.

Anonymous usage signals and any feedback you choose to send remotely also help shape the paid edition. No deal data, financial figures, prompts, or PII are ever collected or used for this. See PRIVACY.md for exactly what is and is not used, and how to keep everything local.

Release Maturity

This release is an internal beta / controlled release candidate. Most top-level skills (1:1 with a single src/skills/<slug>/SKILL.md) are self-contained and usable today. The residential multifamily subsystem and several install surfaces are held to a higher fail-closed bar and are labeled below.

Surface / component Status What this means
Top-level skills (all but one) Deployed Runnable today. Expected behavior per SKILL.md.
residential_multifamily subsystem Stable, awaiting shakedown (v1.0.0-rc1) Code-complete. Routing core, workflows, roles, tailoring guards (approval-floor / canonical-redefinition / preview-bundle / missing-doc blocker), and refusal-on-missing-input contracts are all active and under test. Subsystem still ships with placeholder org overlays; decision-grade use requires an org onboarding pass (tailoring interview) to supply real data. Final-marked outputs fail closed when required inputs are absent, enforce a period-seal gate (close_status, close_lock_timestamp, budget_version), and must follow the executive output contract. Graduation to status: stable is gated on the first operator shakedown log — see PREVIEW_MODE.md for the contract.
Orchestrators Template / semi-manual Orchestrators are phase + agent + verdict templates. There is no autonomous engine that sequences phases, polls agents, or aggregates verdicts without Claude acting as conductor. Treat as structured prompts, not fire-and-forget pipelines.
Marketplace install Supported (CLI) claude plugin marketplace add from Claude Code CLI. Claude Desktop's "Add marketplace" dialog is not supported — this repo does not expose a manifest at that URL. See docs/WHAT-TO-USE-WHEN.md.
macOS DMG / Windows EXE installer Supported Smoke-tested by scripts/installer_smoke_test.py (fresh install).
Cowork ZIP import Partial Skills + agents + commands only. Hooks, MCP tools, orchestrators, calculators are not part of the Cowork surface.
Manual MCP config (Claude Desktop Chat tab) Supported .mcp.json + mcp-server.mjs; operational MCP tools with organizational aliases.
Codex / Gemini / Grok / Manus portable ZIP Experimental (structural CI) Skills ship as SKILL.md files. A structural smoke test (tests/install_smoke/test_portable_zip.py + the Portable ZIP Smoke workflow) validates ZIP layout, skills-tree mirroring, frontmatter contract, and runtime-file exclusion. Cross-runtime invocation (the skill actually loading and running inside Codex / Gemini / Grok / Manus) is not tested. Treat as experimental.

Upgrade, uninstall/reinstall, and corrupted-config recovery paths do not currently have automated smoke tests. The existing smoke tests exercise fresh-install only. See docs/install_smoke_test_matrix.md for the full coverage matrix and gaps.

Known Limitations

Full honest statement: docs/known-limitations.md. The headline v5.2.0 limitations:

  • Connectors are not live. Every connector contract and vendor adapter is status: stub; there is no live API feed to any system. The only active ingestion shape is a shared-drive / email file drop. See docs/connectors/CAPABILITY-MATRIX.md (CoStar is not-supported-live per its AI-use T&C; Yardi / MRI / RealPage are blocked-by-vendor).
  • Generalized cross-skill governance runtime is still deferred. Decision-grade runtime enforcement (source-class tagging, refusal-on-missing-input, period-seal, placeholder scanner) runs as deployed runtime only inside residential_multifamily. Elsewhere the runtime leg is the final_marked selector + the targeted named-allowlist finance-placeholder guard (tests/test_finance_placeholder_guard.py) — a presence-of-discipline check. v5.2.0 adds a corpus-wide static governance scanner (scripts/governance-scan.py) that validates governance declarations over SKILL.md frontmatter + the generated catalog/manifest; it is not a corpus-wide runtime emitted-output scanner. The four canonical connector contract schemas (debt / entity / valuation / funds) remain status: stub (schema-enforced source_class enum), but no connector runtime emits or enforces them and no adapter is live. The live connector runtime and runnable/vendor adapters are v6 work — see Roadmap.
  • The orchestrator runtime dispatchAgent() is a documented stub. The real execution path is the orchestrate prose (src/commands/orchestrate.md) with Claude as conductor — not an autonomous engine. Treat orchestrators as structured prompts.
  • The AMOS manifest is a static export. dist/amos-skill-manifest.json is a generated artifact with honest capability states and no live coupling; no skill is marked live-connected (AMOS references skills as data, it does not invoke them live).
  • The Windows .exe is not locally buildable on macOS — it is produced on CI / Windows only (Inno Setup). The macOS DMG is locally buildable.
  • The residential_multifamily subsystem is status: stable_pending_shakedown (v1.0.0-rc1). Code complete; refusal-on-missing-input contracts active. Every reference file still ships as sample/starter/illustrative/placeholder — decision-grade use requires an org overlay. Final-marked workflows (executive, IC, quarterly, pipeline summary) declare fallback_behavior: refuse on required inputs — they fail closed rather than proceed with stale data — and period-grade workflows additionally refuse if the GL is not at the declared close_status floor (soft_close or hard_close). Graduation to status: stable is gated on the first operator shakedown log (see PREVIEW_MODE.md); until then, output carries a Stable, awaiting shakedown banner.
  • Six regulatory/affordable compliance workflows (compliance_calendar_review, income_certification_cycle, rent_limit_test, agency_reporting_prep, file_audit_prep, recertification_batch) are phase-1 scaffolding — the router recognizes them and the overlay slots exist, but no workflow pack implements them yet. The routing rule r011_regulatory_workflow_explicit is gated behind an explicit regulatory_program axis and rent_limits / income_limits reference files; missing references refuse the match.
  • The tailoring TUI (src/skills/residential_multifamily/tailoring/tools/tailoring_tui.py) has a documented capability matrix in docs/tailoring_capability_matrix.md. Tailoring Pass 2 Obj 4 is closed: conflict surfacing, approval-floor guard, canonical-definition-redefinition refusal, preview-bundle YAML emission, and missing-doc blocker are all Implemented. Open follow-ups (legacy-bank retirement, role-based bank filtering) are tracked in the matrix.
  • Orchestrator pipelines are transitioning from templates to a lightweight runtime. Current /cre-skills:orchestrate invokes the engine for pipelines that declare a config; deal-state persistence, typed approval gates, and a calculator bridge have landed. Verdict aggregation across phases and autonomous challenge-layer resolution remain in progress.
  • Windows installer defends against UTF-8 BOM edge cases in PowerShell 5.1 but does not currently halt on missing Node/Python/npm prerequisites. Update Claude Code before install (older versions have MCP path issues on Windows).
  • Codex / Gemini / Grok / Manus portable ZIP has structural CI coverage (tests/install_smoke/test_portable_zip.py), but cross-runtime invocation is not tested. Treat as experimental on those surfaces.

Roadmap

Upcoming work is tracked in docs/ROADMAP.md. The remaining v5.x work is documentation / contract alignment, static-scanner refinement, manifest capability metadata, no-live invariant enforcement, the generalized runtime governance scanner, and optional non-live interface scaffolding. The real-world / live data-integration track — connector runtime enforcement of source_class / max_staleness, sandbox-runnable adapters, and vendor-specific connectors beyond status: stub — is v6. The enterprise governance track (a lifecycle-hook control plane, four-eyes approvals, deal-state memory, audit, SOC 2, RBAC, licensing, and team collaboration) is the paid cre-skills-pro edition, built separately as the open-core commercial layer on top of this free core; it is not part of this repository's release line. See Open core. Preview / staging mode for status: beta_rc and status: experimental skills is documented at docs/PREVIEW_MODE.md.

Key Stats

Metric Count
Skills 127
Expert Agents 54
Reference Files 279
Python Calculators 21
Workflow Chains 6
Orchestrator Pipelines 10
Slash Commands 11
Skill Categories 18

What's New in v5.1.0

A governance-hardening release — correctness and truthful contracts, not new skills (0 net new). Highlights, each test-backed:

  • Calculator fidelity. Monte Carlo beta-distributed variables now inherit the Gaussian-copula correlation (it was being silently dropped). fund_fee_modeler promote-sensitivity is labeled grade: screening / linear_fee_drag_approximation — the breakeven_irr_* values are a fee-drag approximation, not a DCF-IRR. transfer_tax / proration_calculator / quick_screen now refuse value-domain degeneracy (including the proration unparseable-date path that used to throw a raw traceback). debt_sizing gains interest_only alias keys.
  • Connector contract schemas. The four canonical contracts — debt, entity (legal/ownership, distinct from master_data), valuation, funds (+ pseudonymized investor_report) — are authored as status: stub contracts with round-tripping samples, and source_class is now a schema-enforced enum. No adapter is live; nothing executes.
  • AMOS forward-compat. The skill-manifest emits produces_artifact_kind, pii_policy (conservative default none), and workspace_scope. live_connector stays false.

Still deferred (honest): the corpus-wide runtime governance scanner, the full 127-skill v5_contract sweep, live/runnable connector adapters, and the Windows .exe (CI/Windows-only). See Known Limitations and docs/releases/v5.1.0-release-notes.md.

What's New in v5.0.0

A governable skill taxonomy across the whole corpus. Every catalog item now carries a classification (micro | normal | orchestrator | workspace | agent | calculator) and a runtime_role, so a deal team, asset manager, or fund team can tell an atomic tool from a conductor from a business-facing surface — and, critically, which outputs are decision-grade and need a human sign-off. See How the skill taxonomy works for your team. This shipped by reclassification + honest backlog: 0 new stub skills — the value is the governable metadata and the decomposition map, not skill count.

AMOS skill-manifest export. The plugin now publishes dist/amos-skill-manifest.json (a superset of the catalog carrying the governance fields: decision_grade, human_gate, source_ref_policy, amos_surface, decomposes_to), plus a documented contract at docs/integrations/amos-skill-manifest.md. AMOS consumes the plugin as a governed skill layer as data — no skill is hardcoded, and the manifest is a static generated artifact with no live coupling.

Trust fixes (governance honesty). A canonical data-grade ladder (docs/DATA_GRADES.md) reconciles the four source-class vocabularies already deployed (RMF executive tags, the fallback resolver, the ingestion classification, the connector source_class) into one ladder, and states which grades may back a final-marked output and which must refuse. A targeted finance-placeholder guard (tests/test_finance_placeholder_guard.py) enforces, on a named decision-grade allowlist (acquisition-underwriting-engine, ic-memo-generator, comp-snapshot, fund-lp-reporting, jv-waterfall-architect, opportunity-zone-underwriter, cost-segregation-analyzer), that those skills declare a ## Refusal Behavior discipline and forbid unresolved $X/placeholder tokens in a final output. Valuation/comp skills (comp-snapshot, om-reverse-pricing, acquisition-underwriting-engine) carry an explicit "estimate, not an appraisal" stamp.

Honest scope (read this). This release is a micro-skill architecture delivered as 0 new stubs; reclassify + backlog — not a wave of new atomic skills. The decision-grade governance (source-class tagging, refusal-on-missing-input, period-seal, the placeholder scanner) runs as deployed runtime enforcement only inside residential_multifamily; across the rest of the corpus v5.0.0 ships the final_marked selector plus the named-allowlist finance guard. A fully-generalized, corpus-wide runtime data scanner remains deferred (v5.2.0 later shipped a corpus-wide static declaration scanner; the runtime scanner is still future work). The four canonical connector contract schemas (debt / entity / valuation / funds) landed in v5.1.0 as status: stub; the live connector runtime is a v6 item. See Known Limitations and docs/connectors/CAPABILITY-MATRIX.md — no connector is live today.

See CHANGELOG.md and docs/MIGRATION.md for the full v4.5.0 → v5.0.0 entry.

What's New in v4.3.0

Residential multifamily graduates to stable_pending_shakedown: the subsystem moves from status: beta_rc (v0.6.0) to status: stable_pending_shakedown (v1.0.0-rc1). Tailoring Pass 2 Objective 4 closes — approval-floor guard, canonical-redefinition refusal, preview-bundle YAML emission, and missing-doc blocker are all active (previously listed as Not implemented). Code-complete; awaiting the first operator shakedown log before graduation to status: stable. See docs/PREVIEW_MODE.md for the contract.

Portable ZIP structural smoke: tests/install_smoke/test_portable_zip.py and the Portable ZIP Smoke workflow validate ZIP layout, skills-tree mirroring, frontmatter contract, and runtime-file exclusion. Cross-runtime invocation remains an explicit gap.

Canonical Desktop marketplace caveat: the "Chat tab's 'Add marketplace' is not supported" guidance is now pinned verbatim across README.md, docs/INSTALL.md, docs/install-guide.md, docs/install-desktop.md, docs/install-cowork.md, with a parity assertion in tests/test_release_version_parity.py. docs/WHAT-TO-USE-WHEN.md is the source of truth.

Preview-mode gate is live repo-wide: tests/test_preview_mode_gate.py walks every SKILL.md under src/skills/ and enforces the Release maturity section + banner for each preview sub-status (experimental, beta_rc, stable_pending_shakedown). The residential_multifamily subsystem has its own stricter counterpart under the subsystem tests.

See CHANGELOG.md for the full v4.3.0 entry.

What's Landed in v4.2.0

Hardening pass 2 close: residential_multifamily subsystem moves from status: draft (v0.5.0) to status: beta_rc (v0.6.0). Three deferred objectives close: sealed-close gating (Obj 5) pins period-grade workflows behind a declared close_status floor; a finance-critical placeholder scanner (Obj 6) rejects un-labeled TBD/PLACEHOLDER rows in every CSV read by a final-marked workflow; an executive output contract (Obj 8) requires verdict-first structure + source-class labels on every numeric cell. +13 tests (total 436). See CHANGELOG.md for the full v4.2.0 entry.

Catalog-driven architecture: Single source of truth (src/catalog/catalog.yaml). Every public surface -- README stats, plugin.json, hooks prompt, routing table, registry -- is generated from the catalog. CI catches drift.

MCP server: Zero-dependency MCP server (src/mcp-server.mjs) for Claude Desktop support. macOS DMG and Windows .exe installers auto-detect Claude Code and Claude Desktop.

Workspace skills: deal-intake, lease-strategy-papering, asset-ops-cockpit, capital-projects-development, fund-lp-reporting, navigator, plugin-admin. Persistent workspace state for cross-session continuity.

Feedback system: /cre-skills:send-feedback and /cre-skills:report-problem with automatic redaction, optional remote submission (ask_each_time mode), and retry outbox for failed sends.

Current counts live in Key Stats above; it is regenerated from the catalog by CI. See CHANGELOG.md for full history.

Installation

This repo is a self-contained Claude marketplace. Choose the install method that fits your setup.

See docs/WHAT-TO-USE-WHEN.md for a detailed comparison of all install surfaces.

Do not paste this repo URL into Claude Desktop Chat tab's "Add marketplace" dialog. Chat tab's "Add marketplace" is a separate surface and is not supported by this repo — pasting https://github.com/mariourquia/cre-skills-plugin there will produce a validation error. The canonical Chat tab install path is the DMG (macOS) or EXE (Windows) installer, which registers a local MCP server via claude_desktop_config.json. The Claude Code CLI marketplace (claude plugin marketplace add mariourquia/cre-skills-plugin followed by claude plugin install cre-skills@cre-skills) is supported and is the canonical CLI install path; it also works in the Desktop Code tab (which uses Claude Code under the hood).

Method Best for What you get
Marketplace install Claude Code users (CLI or Desktop Code tab) Full: skills, agents, commands, hooks, MCP tools
macOS DMG Non-technical macOS users Full plugin + Claude Desktop MCP registration
Windows EXE Non-technical Windows users Full plugin + Claude Desktop MCP registration
Cowork upload Cowork tab users Skills, agents, commands (no hooks/orchestrators)
Manual config Chat tab only (MCP tools) MCP tools via claude_desktop_config.json

Marketplace Install (recommended)

From Claude Code CLI or the Desktop Code tab:

claude plugin marketplace add mariourquia/cre-skills-plugin
claude plugin install cre-skills@cre-skills

Or interactively: type /plugin in Claude Code, select Add plugin, and search for cre-skills.

This gives you the full set of skills, agents, commands, hooks, and MCP tools.

Installer (macOS DMG / Windows EXE)

Download from the latest GitHub release:

  • macOS: Download the .dmg, open it, double-click CRE Skills Installer
  • Windows: Download the .exe, run the wizard (SmartScreen: click "More info" > "Run anyway")

The installer auto-detects Claude Desktop, Claude Code, or both and configures each. No admin privileges required. Restart Claude Desktop after installation.

Cowork Tab

In the Claude Desktop Cowork tab, click Customize > Browse plugins to find CRE Skills, or upload the plugin ZIP from the release page. Skills, agents, and commands work in Cowork. Hooks and orchestrators are Code-tab only.

CLI Alternatives

# Install script (downloads and registers automatically)
curl -fsSL https://raw.githubusercontent.com/mariourquia/cre-skills-plugin/main/scripts/install.sh | bash

# Local development
git clone https://github.com/mariourquia/cre-skills-plugin.git
claude --plugin-dir ./cre-skills-plugin

What to Expect After Installation

Claude Desktop has three tabs. Each provides different capabilities:

Code Tab Chat Tab Cowork Tab
Install method Marketplace or installer Installer (MCP config) Cowork plugin import
Skills Yes, via /cre-skills:* No (MCP tools instead) Yes, via / commands
Agents Yes, auto-invoked No Yes, as sub-agents
Skill routing /cre-skills:cre-route cre_route MCP tool Manual
Orchestrators /cre-skills:orchestrate Not available Not available
Workspace Yes Yes (MCP tools) No
Hooks Yes (session start/stop) No No
Customization Yes Yes (MCP tools) No

Code tab gives the full experience. Chat tab gives MCP tools. Cowork tab gives skills and agents without hooks or orchestrators.

Verify Installation

After restarting, try:

/cre-skills:cre-route screen this deal

Or ask Claude: "What CRE skills do you have?"

For a structural check: bash scripts/verify-install.sh

Install and verify also ensure CLAUDE_CODE_STOP_HOOK_BLOCK_CAP is at least 100 in ~/.claude/settings.json (raise-only -- a higher value you already set is left untouched), so the plugin's long agent loops are not truncated by the Stop-hook default of 9.

Troubleshooting

Skills don't appear (Code tab):

  1. Start a new conversation (SessionStart hook fires at start)
  2. Run /plugin to verify the plugin is enabled
  3. If not listed: claude plugin marketplace add mariourquia/cre-skills-plugin

MCP tools don't appear (Chat tab):

  1. Restart Claude Desktop completely
  2. Check Settings > Developer > MCP Servers for cre-skills
  3. Verify Node.js 18+ is installed (node --version)
  4. Re-run the installer if the MCP entry is missing

Cowork tab:

  1. Click Customize > Browse plugins to check if CRE Skills is installed
  2. If not listed, upload the plugin ZIP from the release page

Windows known issues

Update Claude Code first. Older versions of Claude Code on Windows have a bug where colons in MCP log directory paths cause plugin MCP servers to fail silently (anthropics/claude-code#13679). Update to the latest version before installing:

npm i -g @anthropic-ai/claude-code@latest
# or: irm https://claude.ai/install.ps1 | iex

Cowork marketplace empty on Windows. The Cowork plugin marketplace may not load because plugins.claude.ai does not resolve via DNS on some Windows configurations. This is an Anthropic infrastructure issue (anthropics/claude-code#28853). The plugin still works in the Code tab and CLI.

Plugin installed but not showing. If the installer completed but skills don't appear, verify the plugin is enabled in ~/.claude/settings.json:

{
  "enabledPlugins": {
    "cre-skills@local": true
  }
}

The installer writes this automatically, but some Claude Code versions have a bug where enabledPlugins is not populated (anthropics/claude-code#20661).

See docs/INSTALL.md for detailed per-platform instructions.

Quick Start

Screen a deal in seconds:

/cre-skills:cre-route screen this deal -- 240-unit garden-style multifamily,
Raleigh NC, $42M asking, 2018 vintage, 93% occupied, $2.6M NOI, rents 12% below market

Underwrite with full depth:

/cre-skills:cre-route underwrite this deal

Review loan documents:

/cre-skills:cre-route review these loan docs

See all workflow chains:

/cre-skills:cre-workflows

Browse expert agents:

/cre-skills:cre-agents

Run an orchestrated pipeline:

/cre-skills:orchestrate acquisition

Set up brand guidelines (once):

/cre-skills:brand-config

Skill Categories

By Category

# Category Count Key Skills
01 Deal Screening 2 deal-quick-screen, om-reverse-pricing
02 Underwriting & Analysis 4 acquisition-underwriting-engine, rent-roll-analyzer, sensitivity-stress-test, monte-carlo-return-simulator
03 Deal Structuring 6 loi-offer-builder, psa-redline-strategy, jv-waterfall-architect, 1031-exchange-executor, creative-seller-financing, 1031-pipeline-manager
04 Due Diligence 4 dd-command-center, distressed-acquisition-playbook, title-commitment-reviewer, tenant-credit-analyzer
05 Capital Markets 7 loan-sizing-engine, capital-stack-optimizer, refi-decision-analyzer, mezz-pref-structurer, debt-portfolio-monitor, workout-playbook, term-sheet-builder
06 Market Research 4 submarket-truth-serum, comp-snapshot, supply-demand-forecast, market-cycle-positioner
07 Asset Management 7 annual-budget-engine, property-performance-dashboard, capex-prioritizer, noi-sprint-plan, lease-compliance-auditor, tenant-delinquency-workout, small-operator-pm
08 Leasing 9 tenant-retention-engine, lease-up-war-room, lease-negotiation-analyzer, rent-optimization-planner, leasing-operations-engine, leasing-strategy-marketing-planner, lease-document-factory, lease-option-structurer, lease-trade-out-analyzer
09 Investor Relations 11 ic-memo-generator, quarterly-investor-update, lp-pitch-deck-builder, capital-raise-machine, fund-formation-toolkit, fund-operations-compliance-dashboard, investor-lifecycle-manager, sec-reg-d-compliance, distribution-notice-generator, emerging-manager-evaluator, fund-raise-negotiation-engine
10 Development 7 dev-proforma-engine, land-residual-hbu-analyzer, entitlement-feasibility, construction-budget-gc-analyzer, construction-project-command-center, construction-procurement-contracts-engine, post-close-onboarding-transition
11 Disposition 2 disposition-strategy-engine, disposition-prep-kit
12 Deal Sourcing 1 sourcing-outreach-system
13 Tax & Entity 3 cost-segregation-analyzer, opportunity-zone-underwriter, partnership-allocation-engine
14 ESG & Climate 2 carbon-audit-compliance, climate-risk-assessment
15 Portfolio Strategy 3 portfolio-allocator, performance-attribution, deal-attribution-tracker
16 Daily Operations 27 t12-normalizer, rent-roll-formatter, cam-reconciliation-calculator, debt-covenant-monitor, lease-abstract-extractor, estoppel-certificate-generator, cpi-escalation-calculator, variance-narrative-generator, closing-checklist-tracker, vendor-invoice-validator, property-tax-appeal-analyzer, coi-compliance-checker, work-order-triage, lender-compliance-certificate, stacking-plan-builder, building-systems-maintenance-manager, insurance-risk-manager, compliance-regulatory-response-kit, crisis-special-situations-playbook, property-operations-admin-toolkit, tenant-event-planner, loan-document-reviewer, transfer-document-preparer, funds-flow-calculator, gp-performance-evaluator, fund-terms-comparator, lp-data-request-generator

Workflow Chains

Six end-to-end workflow chains that orchestrate multiple skills in sequence.

# Chain Steps
1 Acquisition Pipeline sourcing -> quick-screen -> om-reverse -> rent-roll-analyzer -> underwriting-engine -> sensitivity -> ic-memo -> loi -> psa-redline -> dd-command-center -> close
2 Capital Stack Assembly underwriting-engine -> loan-sizing -> mezz-pref -> jv-waterfall -> capital-stack-optimizer -> refi-decision
3 Hold Period Management annual-budget -> performance-dashboard -> capex / lease-compliance / delinquency-workout / retention-engine -> noi-sprint
4 Disposition Pipeline performance-dashboard -> disposition-strategy -> [SELL] disposition-prep -> 1031-exchange | [HOLD] refi-decision | [REFI] loan-sizing
5 Development Pipeline land-residual + entitlement -> dev-proforma -> construction-budget -> loan-sizing -> capital-stack -> ic-memo -> lease-up-war-room -> refi-decision
6 Fund Management fund-formation -> pitch-deck -> capital-raise -> portfolio-allocator -> [deploy via acquisition pipeline] -> quarterly-update + performance-attribution

Each chain has a detailed workflow document in src/routing/workflows/ with step-by-step orchestration logic.

Orchestration Engine

The plugin includes a multi-agent orchestration engine (derived from Avi Hacker's CRE Acquisition Orchestrator) that coordinates skills into automated pipelines.

Available Orchestrators

Orchestrator Purpose Phases Verdict
acquisition Full acquisition lifecycle DD -> UW -> Financing -> Legal -> Closing -> Challenge GO / CONDITIONAL / NO-GO
capital-stack Optimal debt/equity structuring Qualification -> Sizing -> Structuring -> Optimization -> IC PROCEED / RESTRUCTURE / KILL
hold-period Asset management (recurring loop) Onboarding -> Monitoring (loop) -> Leasing -> Capital -> Tenant -> Reposition CONTINUE / INTERVENE / EXIT
disposition Asset sale pipeline Hold/Sell -> Pricing -> Marketing -> Buyers -> Offers -> DD Mgmt -> Close SELL / HOLD / REFI
development Ground-up development Land -> Entitlement -> Proforma -> Construction -> Lease-Up -> Stabilization BUILD / KILL / DEFER
fund-management Full fund lifecycle Formation -> Raise -> Deploy -> Monitor -> Distribute -> Exit DEPLOY / HOLD / WIND-DOWN
research Market intelligence Macro -> Submarket -> Competitive -> Opportunity -> Memo INVEST / MONITOR / PASS
strategy Investment strategy formulation Capital -> Cycle -> Strategy -> Portfolio -> Memo DEPLOY / REVISE / HOLD
portfolio Portfolio-level oversight Composition -> Concentration -> Attribution -> Rebalance -> Stress -> Report REBALANCE / HOLD / DIVEST
lp-intelligence LP evaluation of GPs GP Eval -> Data Request -> Performance -> Portfolio -> Re-Up RE-UP / REDUCE / EXIT

Run with: /cre-skills:orchestrate acquisition

See src/orchestrators/README.md for full documentation.

Expert Agents

Expert agents across multiple categories, each with a distinct analytical perspective.

Category Count Agents
Institutional Buyers 5 Pension Fund, Private Equity, Public REIT, Family Office, Syndicator
Analytical Lenses 5 Quantitative, Qualitative, Contrarian, Risk Manager, ESG/Impact
Investment Functions 8 Acquisitions Analyst, Asset Manager, Property Manager, Capital Markets, IR Director, Development Manager, Leasing Director, Disposition Strategist
Challenge Agents 6 Conservative Lender, Aggressive GP, Skeptical LP, IC Challenger, Value-Add Operator, Distressed Specialist
Titan Styles 6 Zell, Linneman, Sternlicht, Ross, Gray, Barrack
Stakeholder Views 8 Tenant, Lender, Municipality, Insurance, Appraiser, Environmental, Construction, Legal
Research & Strategy 4 Market Research Analyst, Submarket Specialist, Chief Investment Officer, Portfolio Strategist
Asset Management 2 Asset Manager Lead, Leasing Manager
Portfolio 2 Portfolio Manager, Risk Officer
Fund Management 2 Fund Controller, Investor Relations Associate
Disposition 1 Disposition Manager
LP Intelligence 3 LP Advisor, Fund Analyst, Allocation Committee Member
Composite 2 CRE Veteran (generalist router), Deal Team Lead (multi-agent orchestrator)

The Deal Team Lead agent assembles multi-agent teams from 10 pre-built compositions: Acquisition IC, Capital Stack Optimization, Disposition Strategy, Development Feasibility, Lease Negotiation, Fund Formation, Market Cycle Assessment, Crisis Response, Portfolio Review, and LP Due Diligence.

Python Calculators

21 standalone Python scripts (zero external dependencies) that agents can execute for precise quantitative output. The table below lists the most-used; the full set is in src/calculators/.

Script Skill Calculations
quick_screen.py deal-quick-screen Cap rate, DSCR, CoC, replacement ratio, 3-scenario IRR
debt_sizing.py loan-sizing-engine DSCR/LTV/debt yield constraint optimization
covenant_tester.py loan-document-reviewer DSCR/LTV/debt yield by year, breach detection
npv_trade_out.py lease-trade-out-analyzer NPV comparison, breakeven, 2D sensitivity grid
option_valuation.py lease-option-structurer Termination fees, cap rate impact, package NPV
waterfall_calculator.py jv-waterfall-architect GP/LP distributions, multi-tier promote, IRR
tenant_credit_scorer.py tenant-credit-analyzer HHI, WALT, expected annual loss, OCR
proration_calculator.py funds-flow-calculator Per diem prorations (actual/365, 30/360)
transfer_tax.py transfer-document-preparer All 50 states + DC with tiered rates
monte_carlo_simulator.py monte-carlo-return-simulator Stochastic return distributions, confidence intervals, tail risk
fund_fee_modeler.py fund-raise-negotiation-engine Management fee modeling, carried interest schedules, blended fee load, MFN cascade

All calculators are in src/calculators/. Run directly with python3 src/calculators/<script>.py.

Brand Guidelines

Skills that produce investor-facing deliverables (pitch decks, IC memos, investor updates, offering packages, leasing marketing materials) automatically load your brand guidelines from ~/.cre-skills/brand-guidelines.json.

Run /cre-skills:brand-config to set up your brand colors, fonts, disclaimers, and contact info once. All future deliverables will use your brand automatically.

What Gets Saved

Setting Description
Company/fund name Appears in all headers and footers
Primary, secondary, accent colors Applied to formatting instructions
Heading and body fonts Referenced in all layout directives
Layout style minimal, corporate, boutique, or institutional
Number format full ($1,234,567), abbreviated ($1.2M), or both
Units preference psf, per_unit, or auto
Disclaimer text Appended to every page/section
Confidentiality notice Applied to cover pages
Contact information block Placed on final page of deliverables
Logo file path Referenced (not embedded)

Skills That Auto-Load Brand Guidelines

Skill Deliverable Type
lp-pitch-deck-builder LP pitch decks (16-slide)
ic-memo-generator Investment committee memos
quarterly-investor-update Quarterly LP letters and reports
capital-raise-machine LP packs, capital call notices, onboarding materials
fund-formation-toolkit PPM drafting guidance, fund term materials
disposition-prep-kit Offering packages, buyer marketing materials
investor-lifecycle-manager LP meeting prep, benchmark reports, GIPS composites
leasing-strategy-marketing-planner Leasing flyers, brochures, marketing plans

First-Use Behavior

On first invocation of any deliverable skill, if no brand guidelines file exists, the skill will prompt:

"I don't have your brand guidelines saved yet. Would you like to set them up now with /cre-skills:brand-config? Or I can proceed with professional defaults."

Professional defaults: navy #1B365D, white #FFFFFF, gold accent #C9A84C, Helvetica Neue/Arial fonts, standard CRE disclaimer language.

Feedback

Share feedback or report problems without leaving your session:

  • /cre-skills:send-feedback -- share feedback about skill quality, missing capabilities, or general suggestions
  • /cre-skills:report-problem -- report a bug with structured severity, reproduction context, and skill identification

Feedback is saved locally to ~/.cre-skills/feedback-log.jsonl. Free-text fields are automatically sanitized (file paths, emails, digit sequences stripped). No deal data, prompts, or financial figures are ever stored. After submitting, you're asked if you'd also like to send it to the maintainer (you approve each send). To disable: set feedback.mode to local_only in ~/.cre-skills/config.json.

View your feedback history with /cre-skills:feedback-summary.

Skill Customization

Adapt any skill to how your team actually works. Local overrides take priority over base skills -- the base files are never modified.

/cre-skills:customize-skill

The plugin walks you through: select a skill, make changes, record why. Common customizations:

Category Example
Terminology Rename fields to match your organization
Approval chain Add compliance officer review steps
Required steps Insert ESG screening into underwriting
Deliverable format Restructure IC memo for your committee
Calculation method Use MOIC instead of IRR as primary metric
Regional / market Add NYC transfer tax tiers

Customizations are stored at ~/.cre-skills/customizations/<slug>/ and persist across plugin updates. You can optionally share structured feedback about your changes with the maintainer to help improve the plugin.

Privacy: Default mode is metadata_only -- only skill name, change categories, and rationale are shared. No skill content leaves your machine without explicit consent. Set customization.feedback_mode to off in ~/.cre-skills/config.json to disable entirely.

See docs/customization-guide.md for full details, configuration examples, and MCP tool reference.

Privacy & Telemetry

Anonymous usage telemetry is enabled by default and local-only. It records which skills you use (slug only) and the date -- nothing else. No deal data, financial figures, file paths, prompts, or identity information is ever tracked. All data stays on your machine in ~/.cre-skills/telemetry.jsonl. To opt out: set "telemetry": false in ~/.cre-skills/config.json.

  • /cre-skills:usage-stats -- view your aggregated skill usage patterns
  • /cre-skills:feedback-summary -- view your session ratings and comments
  • /cre-skills:send-feedback -- share structured feedback (saved locally)
  • /cre-skills:report-problem -- report bugs (saved locally)

See PRIVACY.md for what is and is not collected.

Project Structure

cre-skills-plugin/
  src/
    plugin/
      plugin.json            # Plugin manifest
    skills/
      <slug>/
        SKILL.md             # Skill definition (process, inputs, outputs)
        references/          # Supporting reference documents (.md and .yaml)
    agents/
      _index.md              # Agent roster and team compositions
      <agent>.md             # expert agent definitions (flat directory)
    orchestrators/
      engine/                # Pipeline engine schema and handoff registry
      configs/               # orchestrator JSON configurations
      prompts/               # orchestrator prompt files
      challenge-layer/       # Post-pipeline adversarial review config
      investor-profiles/     # 8 investor profiles + strategy matrix
      schemas/               # Disagreement and reversal trigger schemas
      thresholds.json        # Investment thresholds (base + investor overrides)
      README.md              # Orchestration engine documentation
    commands/
      cre-route.md           # Skill router command
      cre-workflows.md       # Workflow chain browser
      cre-agents.md          # Agent roster browser
      brand-config.md        # Brand guidelines setup
      customize-skill.md     # Skill customization workflow
      orchestrate.md         # Multi-agent pipeline orchestrator
      usage-stats.md         # Telemetry summary
      feedback-summary.md    # Session feedback log
      send-feedback.md       # Share feedback
      report-problem.md      # Report a bug
    lib/
      customization.mjs      # Skill override CRUD and resolution
      diff.mjs               # LCS-based line diff engine
      feedback-payload.mjs   # Customization feedback payload builder
    routing/
      CRE-ROUTING.md         # Master routing index
      workflows/             # Detailed workflow chain documents
    hooks/
      hooks.json             # Hook definitions (SessionStart, PostToolUse, Stop)
      telemetry-init.mjs     # Initializes user config on first run
      telemetry-capture.mjs  # Tracks skill invocations (opt-in)
      session-summary.mjs    # Session end record and feedback
    calculators/             # 12 Python calculator scripts
    catalog/                 # Catalog schema and canonical catalog.yaml
    mcp-server.mjs           # MCP server for Claude Desktop
    templates/
      output-styles/         # Output format templates
  scripts/
    redact-feedback.mjs    # Feedback sanitization utility
    install.sh             # Fresh install with v1->v2 migration
    update.sh              # Pull latest, detect breaking changes
    uninstall.sh           # Clean removal with data preservation
    verify-install.sh      # 8-check health report
    ensure_stop_hook_cap.py # Raise CLAUDE_CODE_STOP_HOOK_BLOCK_CAP to >=100 (raise-only)
  registry.yaml            # Skill registry with metadata and chain mappings
  tests/                   # pytest suite: structural integrity, catalog parity,
                           # release hygiene, docs/surface/version parity, MCP server,
                           # orchestrator engine (deal state, gates, calculator bridge),
                           # installer hardening, and end-to-end calculator exec

Attribution

This release integrates orchestration patterns derived from the CRE Acquisition Orchestrator by Avi Hacker (The AI Consulting Network), licensed under Apache 2.0. See NOTICE for full attribution.

Contributing

See CONTRIBUTING.md for guidelines on adding new skills, agents, and calculators.

License

Apache 2.0

See NOTICE for attribution and patent grant details.

Author

Mario Urquia -- mariourquia.com

Quant, Data, and Startup Product Builder. Building proof-of-concepts for commercial real estate investment management across equity, debt, and infrastructure.

本页内容由上游仓库自动镜像，如有疑问请访问原始仓库。