🐦⬛ RavenClaude
A public Claude Code plugin marketplace — governed multi-agent teams, specialist plugins, and a comfort-posture dashboard that travel with you across projects.
🚀 ▶ See what RavenClaude is (pitch.html) — value prop, proof, and catalog at a glance. Start here.
📖 ▶ Open the RavenClaude portal — browse every plugin, agent, skill, hook, rule, and template (with an “I want to…” lookup). Regenerated from the manifests on every release.
🎛 ▶ Open the RavenClaude dashboard — comfort-posture editor (preview). For a live Save & apply against your repo, run rc dashboard (or bash scripts/open-dashboard.sh in this checkout).
📖 ▶ First Workflow in 10 Minutes — install → dashboard → one governed multi-agent dispatch → /wrap.
🌐 ▶ Raven Power ↗ — the consulting front door behind RavenClaude.
Today this marketplace ships 184 plugins (live inventory: portal Marketplace — prefer that over any hand-maintained essay list below).
ravenclaude-core is the foundation: Team Lead + specialist agents, dispatch playbooks, gates, hooks, templates, comfort-posture, and the Learn / teaching surfaces on the dashboard. Domain plugins (Power Platform, cloud, finance, security, and many more) install beside it.
What's new in core (through 0.324.7)
- Thing tribunal — opt-in command-review for risky tools; hardening EDIT behind
command_review.hardening_edit (default ON; set false to opt out). Force-push / mirror-push / hard-rule floors stay DENY under EDIT ON.
- Runes ready-queue — Oath-hook +
rc runes; dashboard Runes at session start opt-in (runes: absent ⇒ off); may auto-claim ungated ready work when On; never auto Longship merge.
- Prompt Builder —
#/prompt-builder with Gate 144 / Host Context floors; catalog model SSOT; blank modes + canned Task/Few-shot.
- Plugin lifecycle — per-project last-used ledger;
auto_uninstall / auto_install default OFF (auto_install: off|ask|auto); opt-in uninstall execute when ON; tip/SHA integrity pin required before auto install --execute (fail-closed on missing/mismatch); ask-first install from the ravenclaude marketplace only; ravenclaude-core hard-pinned.
- Blocked-exhaustion gate — after a RavenClaude guard denies a tool, hand-backs to the human are refused until distinct-channel workarounds are logged (
rc workaround …); knob workaround_exhaustion: off|warn|block (absent ⇒ off).
- Comfort-posture + dashboard — deny/ask/allow autonomy; Settings + Pipeline; host-scope badges; Learn tab for teaching the system.
Count-drift note: specialist/skill/hook tallies change every release. Treat the portal and claude plugin details ravenclaude-core as SSOT — do not trust stale “15 specialists / 58 skills” prose if it reappears elsewhere. Marketplace plugin count on tip: see portal / .claude-plugin/marketplace.json (currently 184).
Install
This repository is public: anyone can read it, clone it, and install from it without being added as a collaborator. Write access is limited to the collaborators listed in docs/access.md.
Quick start in Claude Code (inside a session):
/plugin marketplace add mcorbett51090/RavenClaude
/plugin install ravenclaude-core@ravenclaude
/reload-plugins
Or from your shell:
claude plugin marketplace add mcorbett51090/RavenClaude
claude plugin install ravenclaude-core@ravenclaude
The marketplace name is ravenclaude (from .claude-plugin/marketplace.json), so every install id is <plugin>@ravenclaude.
There are two install paths, depending on which agent host you use. Pick one and follow that column — the surfaces don't overlap (the /plugin slash commands exist only in Claude Code; the ravenclaude script exists only on the Copilot side).
| Path A — Claude Code (recommended if you have it) | Path B — GitHub Copilot CLI / Codespace |
|---|
| Audience | You run claude and want the marketplace's agents/skills/hooks inside it. | You run copilot (or want the dashboard + governance in a Codespace without Claude Code). |
| Install | /plugin marketplace add mcorbett51090/RavenClaude /plugin install ravenclaude-core@ravenclaude /reload-plugins | git clone https://github.com/mcorbett51090/RavenClaude.git ~/RavenClaude bash ~/RavenClaude/scripts/ravenclaude setup --project . source ~/.bashrc |
| Add a plugin | /plugin install power-platform@ravenclaude /reload-plugins | bash ~/RavenClaude/scripts/ravenclaude setup --project . --with-plugin power-platform |
| Update later | /plugin marketplace update ravenclaude /reload-plugins | rc (the alias = ravenclaude update && copilot --plugin-dir …) |
| Pin a SHA | /plugin marketplace add mcorbett51090/RavenClaude#<sha> | git -C ~/RavenClaude checkout <sha> |
| Launch | (nothing — Claude Code loads the plugin automatically) | rc in a NEW terminal (or bash -i -c rc from a non-interactive shell) |
That's it. The ravenclaude-core specialist agents become available to the Team Lead via the spawn-team skill, the dispatch skills are loaded, and the format/lint/test hooks fire automatically. Installing a domain plugin (e.g. power-platform) adds its specialists alongside.
Path B — zero-touch Codespace auto-setup
For a brand-new repo you can skip even those commands. Stamp the Codespace template into the repo once:
bash ~/RavenClaude/scripts/ravenclaude init-codespace --project /path/to/repo
It drops .devcontainer/devcontainer.json + .devcontainer/ravenclaude-post-create.sh into the repo (or, if a devcontainer.json already exists, names the keys to merge — postCreateCommand, postStartCommand, forwardPorts, portsAttributes). Commit those, rebuild the Codespace, and on every start the post-create script installs prerequisites (Node 22+, git-lfs, Copilot CLI), wires the repo, applies the balanced posture, and adds the rc alias. The dashboard auto-launches on the forwarded port — no command to remember.
Path B prerequisites (most images already have them)
The post-create script auto-installs these on Debian-family images, but if you're on an image without apt-get you need them present yourself:
- Node 22+ (Copilot CLI requires it). The template image
mcr.microsoft.com/devcontainers/universal:2-linux already has Node; the python:3.12 family does not.
- git-lfs (for any repo with LFS-tracked assets).
- GitHub CLI (
gh) — used to clone the marketplace (or your own private fork of it) via your Codespace's auth.
The three dashboards (don't confuse them)
RavenClaude ships three dashboard surfaces with similar URLs but different scopes. Most install confusion comes from clicking the wrong one.
| Surface | URL / launch | What it edits | When to use |
|---|
| Published preview (read-only) | https://mcorbett51090.github.io/RavenClaude/plugins/ravenclaude-core/dashboard.html | Nothing — Save & apply is a no-op (no server) | Browse the UI before installing. Do not try to set a real posture here. |
| Marketplace local dashboard | bash scripts/open-dashboard.sh (from a marketplace clone) | This marketplace repo's own .ravenclaude/comfort-posture.yaml | Only when developing the marketplace itself (you're inside RavenClaude/). |
| Per-repo consumer dashboard | bash .ravenclaude/dashboard.sh or ravenclaude dashboard (auto-launches on Codespace start if init-codespace ran) | Your repo's .ravenclaude/comfort-posture.yaml + .claude/settings.json | Every other case. This is the one consumers should use. |
If Save & apply seems to do nothing, you're almost certainly on the first row. Switch to the second or third.
Local development install
If you want to iterate on this marketplace itself (or test a change before pushing):
# From any test project, point at your local checkout:
/plugin marketplace add /path/to/RavenClaude
/plugin install ravenclaude-core@ravenclaude
/reload-plugins
After editing files in plugins/ravenclaude-core/ (or plugins/power-platform/), run /plugin marketplace update ravenclaude and /reload-plugins again to pick up the changes.
Fallback — clone instead of install
If you're on a Claude Code plan or in an enterprise environment that restricts marketplace installs, you can clone this repo and copy the plugin folder manually:
git clone https://github.com/mcorbett51090/RavenClaude.git
# User scope (available across all your projects on this machine):
cp -r RavenClaude/plugins/ravenclaude-core/* ~/.claude/
# Or project scope (just one project):
cp -r RavenClaude/plugins/ravenclaude-core/* /path/to/your/project/.claude/
You lose auto-update and version pinning. To update, git pull and re-copy. Otherwise the agents, skills, hooks, rules, and templates work identically.
Updating and version pinning
The marketplace ships semver-versioned plugins (plugin.json version + matching marketplace.json entry, CI-gated for drift). 183 of the 184 plugins declare requires.ravenclaude-core — a minimum ravenclaude-core version they expect, surfaced in the per-plugin card of the portal’s Marketplace section (index.html).
To update everything to the marketplace's latest:
/plugin marketplace update ravenclaude
/reload-plugins
That pulls the catalog head + reloads every installed plugin. Safe for day-to-day use; CI's version-drift gate catches manifest mismatches before they ship.
To pin to a specific commit SHA (recommended for client engagements where surprise updates are unwelcome):
/plugin marketplace add mcorbett51090/RavenClaude#<git-sha>
/plugin install ravenclaude-core@ravenclaude
/reload-plugins
The pin survives /plugin marketplace update — the pinned SHA is the catalog's source of truth for that engagement until you re-add at a newer SHA.
To check compatibility between a domain plugin and your installed ravenclaude-core: open the portal (index.html) → Marketplace, find the plugin, read the Requires row. If your installed core version is older, update core first (/plugin install ravenclaude-core@ravenclaude to latest, or pin to a SHA ≥ the requirement).
When an upgrade prompts an ask in the comfort-posture dashboard: that's expected — shell_package_install defaults to ask in the balanced seed (added v0.101.0). Click Allow once the first time; flip the category to allow from the dashboard's Set up tab if you'd rather not see the prompt.
The non-removable security floor (force-push, rm -rf, curl | sh, host credential reads — ~/.ssh, ~/.aws, etc.) cannot be wiped by editing comfort-posture.yaml — apply-comfort-posture.py always unions the baseline with whatever the user supplies. Verified by tests/fixtures/test_security_deny_floor.py. See SECURITY.md §"Defaults and floors" for the full list.
What's in the marketplace
184 plugins (verified). Full inventory with agents, skills, hooks, and “I want to…” lookup: portal Marketplace. Per-plugin detail lives in each plugins/<name>/README.md.
Featured (starter set)
| Plugin | Focus |
|---|
ravenclaude-core | Domain-neutral Claude Code foundation: specialist agents, skills, gates, hooks, rules,… |
power-platform | Power Platform specialist team — agents (incl. power-platform-tester, power-bi-engineer) and… |
claude-app-engineering | Claude app-engineering specialist team — agents (claude-solution-architect,… |
security-engineering | Security-engineering (AppSec) team — agents (appsec-engineer, threat-modeler,… |
incident-response-dfir | Blue-team DFIR / SOC team — agents (dfir-response-lead, detection-and-forensics-engineer) for… |
aws-cloud | AWS cloud infrastructure & platform team — agents (aws-architect, aws-iam-identity-engineer,… |
azure-cloud | Azure cloud infrastructure & platform specialist team — agents (azure-architect,… |
finance | Corporate finance & FP&A specialist team — FP&A analyst (budgets, forecasts, variance… |
|
Install any plugin with /plugin install <name>@ravenclaude then /reload-plugins (Path A), or ravenclaude setup --project . --with-plugin <name> (Path B).
Contributing back from a consumer project (no repo access needed)
You don't need write access to this marketplace to propose a lesson back. If you're working in any consumer project that has ravenclaude-core installed and Claude discovers a pattern, fix, or rule worth keeping, use the contribution-staging loop:
- In your consumer session, ask Claude to use the
contribute-finding skill on the finding. It formats a canonical RAVENCLAUDE-STAGING-SUBMISSION block (lesson or best-practice shape).
- Send the block to Matt (Slack, email, paste in a shared doc).
- On the marketplace side, Matt drops the block into
docs/staging/incoming/ and runs /review-staged-contributions — security sweep + topic-expert routing, then keep / update / deny.
Full flow: docs/staging/README.md. This is the design-intent contribution path for collaborators who don't (or shouldn't) need direct push access.
How agents actually get invoked
A common point of confusion: these plugin agents do not appear as subagent_type options on the Agent tool. Claude Code's built-in agent list (general-purpose, Explore, Plan, etc.) is separate from plugin-supplied agents, and a fresh Agent call with subagent_type: "code-reviewer" will fail with an InputValidationError.
Plugin agents fire through the Team Lead orchestration pattern:
- The top-level Claude session acts as Team Lead — it reads
plugins/ravenclaude-core/CLAUDE.md, sees the team roster, and decides which specialist(s) the user's request needs.
- To dispatch one or more specialists, the Team Lead invokes the
spawn-team skill — that's the playbook for picking the right specialist, briefing it like a new colleague, and integrating the structured handoff payload that comes back.
- The specialist runs, returns a Markdown report ending in a
---RESULT_START--- … ---RESULT_END--- JSON block, and the Team Lead re-routes from there.
- Sub-agents never spawn other sub-agents. They return a slice; the Team Lead re-dispatches. This keeps the dependency graph a flat tree.
If you want to talk to a specific agent directly (e.g. "have the architect look at this"), say so in plain English to the Team Lead and it will use spawn-team to dispatch the architect. Don't try to address the agent by name through the Agent tool's subagent_type parameter — that's reserved for the built-in agents.
For the dispatch playbook itself, see plugins/ravenclaude-core/skills/spawn-team/SKILL.md.
Browsing the marketplace at a glance
index.html at the repo root is the portal — an interactive single page covering every plugin, agent, skill, hook, rule, and template that ships from this marketplace (in the Marketplace section), plus the comfort-posture Dashboard. Open it in any browser (no server required) for a searchable view of the current state — or click ▶ Open on GitHub Pages to render it from main without cloning. It is regenerated from the manifests on every release via python3 scripts/generate-index-dashboard.py; CI's freshness gate (Gate 97) fails if it drifts.
Working on the marketplace itself
If you're developing RavenClaude (adding plugins, updating agents), see CLAUDE.md at this repo's root — it's the meta-repo dev guide.
Repo layout:
RavenClaude/
├── .claude-plugin/marketplace.json ← marketplace catalog
├── plugins/
│ ├── ravenclaude-core/ ← the domain-neutral plugin
│ └── power-platform/ ← Microsoft Power Platform specialists
├── .claude/ ← settings for working ON the marketplace
├── docs/ ← meta-repo docs
├── checklists/ ← release / new-plugin / incident checklists
└── CLAUDE.md ← meta-repo dev guide
The container at .devcontainer/ auto-installs the Claude Code CLI on rebuild, so a fresh Codespace is ready to work on plugins without setup.
Claude plugin Directory limits — keep the repo submittable
Anthropic's developer portal checks a plugin before listing it in the Claude plugin directory. The size limits are the ones this repo can break just by growing. Source of truth: the pre-submission checklist, re-read 2026-10-14. Re-read it before trusting any number below; they drift.
| Limit (the checklist's wording) | If exceeded | Monorepo / full plugin now | Edition now | Checked by |
|---|
| Repository: "fewer than 10,000 files and folders" | Validation stops | 9,907 files + 3,042 folders = 12,949 | 144 + 41 = 185 | Gate 302 counts files only; --report shows folders |
| Repository: under 50 MiB "as GitHub archives it", 256 MiB unpacked | Validation stops | 41.2 MiB zip, 140.8 MiB unpacked | 1.3 MiB | --report (not gated) |
| Every file in the plugin folder under 5 MiB | Validation stops | dashboard.html, 10.5 MiB | largest 68 KiB | Gate 301 |
| Plugin folder at most 512 files | Held for a reviewer | 895 | 144 | Gate 301 |
| Non-image, non-font file under 256 KiB; only text, SVG, PNG/JPEG/GIF/WebP and fonts | Held for a reviewer | 46 files over | none | Gate 301 |
| File and folder names valid on Windows and macOS; no symlinks, submodules, LFS pointers or system files | Validation stops / blocks | clean (checked 2026-10-14) | clean | not gated |
| README of at least 40 words outside code blocks, and a LICENSE | Blocks | — | both present | Gate 301 |
What is not settled. Two readings are unverified, and only the portal's Validate answers either. First, whether the portal adds files and folders together; the wording reads that way, and on that reading the monorepo is 2,950 over. Second, whether the text eol=lf lines in the root .gitattributes count as the "attributes that rewrite file contents" the checklist says to keep out (the edition ships no .gitattributes). So the plan of record is to submit the directory edition from its own repo (docs/directory-edition/), which is well inside every limit. The monorepo's file count is kept down as insurance, not as the submission path.
Keeping it that way. Before adding many files or a large artifact, run:
python3 scripts/check-repo-file-budget.py --report # files, folders and sizes against every limit (informational)
python3 scripts/check-repo-file-budget.py # Gate 302: fails at 9,990 tracked files
python3 scripts/check-directory-edition.py # Gate 301: the edition against the plugin limits (about 60 s)
If Gate 302 fails, pack generated per-file artifacts into a few files (the pattern is scripts/visual_pack.py) or prune dead files. Don't raise FAIL_AT without a reason in the PR. scripts/audit-gates.sh runs both gates, so the normal pre-PR check covers them.
Roadmap
Shipped since the original roadmap: finance, regulatory-compliance, web-design, edtech-partner-success, data-platform, applied-statistics, microsoft-fabric (the enterprise-Microsoft data-platform lane — OneLake / Lakehouse / Warehouse / Data Factory / Real-Time Intelligence / Direct Lake / capacity FinOps, from docs/microsoft-fabric-plugin-analysis.md), claude-app-engineering (building on the Claude API + Agent SDK + MCP, from docs/claude-app-engineering-plugin-analysis.md), and azure-cloud (Azure infrastructure & platform, from docs/azure-cloud-plugin-analysis.md).
salesforce (Apex, Flow, Agentforce, platform-architecture specialists) has since shipped as well — it is now one of the 184 plugins above, no longer planned-only.
Each builds on top of ravenclaude-core (which provides the neutral team) and adds domain-specific agents that the consumer can choose to install or skip. power-platform is the reference implementation of this pattern.
Security
Found a vulnerability, or a secret that shouldn't be here? Don't open a public issue. Report it privately as described in SECURITY.md.
License
MIT — see LICENSE for the full text. Bundled third-party content carries its own attribution; see plugins/power-platform/NOTICE.md for the Daniel Kerridge skills import and the pbix-mcp server attribution.
