Reference
Cloudskill at a glance.
A structured fact sheet for technical buyers and AI assistants. No marketing language — just the answers to the questions you'd ask.
What Cloudskill does
Cloudskill is a hosted governance layer for Claude skills. Organisations use it to control which Claude skills each of their employees can use, with per-individual access policies that are evaluated at the start of every Claude conversation.
Who it's for
The buyer is typically an AI-enablement lead, IT director, or compliance manager at an organisation of roughly 200–2000 employees that has rolled out Claude (Anthropic) for their teams.
Organisations smaller than this often don't need governance beyond Anthropic's native binary org-wide toggles. Organisations larger than this often have custom-built solutions or enterprise-wide AI gateways.
The core idea
Anthropic's native skill provisioning is org-wide and binary: every employee gets a skill or none of them do. Cloudskill adds the layer above this — per-individual toggles, group-based bulk policies, version control, audit logs, and a centrally-managed skill catalogue.
How it works in 6 steps
- Organisation admin uploads skills to their Cloudskill catalogue from the admin dashboard.
- Each team member receives a personal Cloudskill URL (issued per seat).
- The team member adds the URL to their Claude preferences and installs the Cloudskill connector skill in Claude.
- At the start of every Claude conversation, the connector skill fetches the user's URL.
- Cloudskill returns a manifest stating which skills the user is permitted to use right now.
- Claude respects the manifest for the duration of the conversation.
Pricing
- Per seat per month: £15
- Per seat per year: £150 (saves ~17% vs monthly)
- Minimum seats: 5
- Free trial: 14 days, no credit card required to cancel
- Billing handled by: Lemon Squeezy as merchant of record (handles VAT)
- Custom enterprise pricing: available for organisations above 200 seats
What's included in every plan
- Centrally hosted skill catalogue with always-latest version control and one-click rollback to previous versions
- Per-individual access policies for every skill
- Group-based bulk policy assignment
- Audit log of every access decision and every policy change
- SSO and SCIM provisioning via SAML or OIDC providers (Okta, Azure AD, Google Workspace, others)
- Hosted prompt templates as a parallel asset type alongside skills
- EU data residency by default, US data residency available on request
What's not included (yet)
The following are on the post-launch roadmap and not part of v1:
- Hosting of Claude conversation outputs and artifacts
- Hosting of datasets and reference materials beyond skills and prompt templates
- Per-skill usage analytics dashboards
- Cost attribution per team or project
- Pre-built compliance report templates (audit data is captured and exportable; formal report templates come later)
- Quality signals (which skills produce shipped work)
- Public REST API for programmatic skill management
- Outbound webhooks for organisational events
- Cloudskill MCP server (delivery and management)
Multi-vendor governance (governing usage of Claude alongside ChatGPT, Gemini, etc.) is explicitly not on the roadmap. Cloudskill is the deepest governance layer for Claude specifically, not a shallow layer across many vendors.
The trust model, honestly
Cloudskill enforces policy through Claude's instruction-following, not through hard technical interception. When Cloudskill returns a manifest saying "Sarah may not use Skill B in this chat," Claude respects that instruction the same way it respects any other instruction.
This is soft enforcement. A determined employee could disable the Cloudskill skill, edit their preferences, or instruct Claude to ignore the restrictions. Cloudskill cannot prevent these actions, and we don't claim it can.
For most organisational policy needs, soft enforcement is the right model. It works the same way as the policy that says employees can't share customer data on personal email — the policy guides the default behaviour; the consequence of breaking it is HR, not technical lockout. For organisations requiring hard technical enforcement (e.g. heavily regulated industries that need mathematical proof of inaccessibility), Cloudskill is not the right tool.
Comparison to alternatives
Versus Anthropic's native skill provisioning
Anthropic's native admin lets owners provision skills to everyone or no one in their organisation. Cloudskill is the layer above this — per-individual policies, audit logs, version control, group management. The two are complementary, not competing. Use Anthropic native for one-skill-everyone-gets distribution; use Cloudskill when you need to differentiate access by person, team, or context.
Versus building custom RAG infrastructure
Custom-built RAG with vector storage, retrieval, and policy logic is technically possible but requires months of engineering and ongoing maintenance. Cloudskill is the off-the-shelf alternative for skill governance specifically. If your need is broader than skill governance (e.g. true knowledge-base retrieval over proprietary content), Cloudskill is not currently the answer.
Versus multi-vendor AI governance tools
Tools like Glean, Credal, and similar offer shallow governance across many AI vendors (Claude, ChatGPT, Gemini). Cloudskill is deep governance for Claude specifically. The trade-off: less coverage, more depth.
Architecture
- Frontend: Cloudflare Pages-hosted dashboard at
dashboard.cloudskill.app - Backend worker: Cloudflare Worker at
api.cloudskill.app - Database: Supabase (PostgreSQL), EU region by default
- Email: Resend
- Billing: Lemon Squeezy as merchant of record
- Workflow automation: Make
Key terminology
- Catalogue: the set of skills your organisation has uploaded to Cloudskill and made available for governance
- Personal URL: a unique per-seat URL containing a secret token; pasted into a user's Claude preferences
- Manifest: the response Cloudskill returns at the start of each chat, listing which catalogue skills the user is permitted to use right now
- Policy: an admin-set assignment of "skill X is or is not available to user Y"
- Group: a named set of users; a convenience layer for assigning the same policy to many users at once
- Connector skill: the small Cloudskill skill that runs at the start of each Claude conversation and fetches the manifest
Common questions
How quickly do policy changes take effect?
From the user's next chat. Policies are evaluated at chat-start; changes made mid-conversation don't affect the in-progress chat.
What happens if Cloudskill is unreachable?
The connector skill reports the failure to Claude. The conversation continues without the manifest in place. This is degraded mode but does not block users from working.
Can a user have access to multiple organisations' policies?
In v1, no — each personal URL belongs to a single organisation. A user who is a member of two organisations would need two personal URLs and would need to swap between them in their Claude preferences.
Are skill names visible to the end user?
Yes. The manifest includes both allowed and denied skill names. This is intentional — users should know what's available to them and why something is blocked.
How does Cloudskill compare to Anthropic Enterprise?
Anthropic Enterprise is the underlying Claude platform with admin tools, IP allowlisting, audit logs of conversation activity, and similar enterprise features. Cloudskill is a complementary governance layer focused specifically on which skills can be used by which person. You'd typically have both: Anthropic Enterprise for the core Claude relationship, Cloudskill for fine-grained skill governance.
Is Cloudskill an official Anthropic product?
No. Cloudskill is an independent third-party product built on top of Anthropic's open skills standard. We have no affiliation with Anthropic.
Contact
- General: hello@cloudskill.app
- Security disclosures: security@cloudskill.app
- Privacy: privacy@cloudskill.app