Keep the one-command install; put a versioned, governed marketplace behind it
Asked (summary):
We centralize our OpenCode agent skills, tools, plugins and agents in one repository, installed via a homemade npx @company/agent-skills installer from Artifactory that copies files into user paths. What is the state of the art — how do other companies share and centralize their agent-harness tooling, and make install easy even for non-power users?
Evidence: 20 researched rows — 10 native platform mechanisms (OpenAI Codex/ChatGPT plugins, Claude Code, GitHub Copilot, the open Agent Skills standard) and 10 company implementations (Cloudflare, Expedia Group, Cisco, Tencent and others) — distilled from 3,454 candidate documents across 15 independent hosts, as of 2026-09-24. Seven company rows rest on secondary sources and are marked as such throughout.
Target architecture: the npx command becomes a thin client for a control plane
supply chain you already trustnew control planedeveloper machinecentrally operated tools
Hover or tap a box for what it does and which evidence rows use the same pattern.
What the evidence says
Cloudflare onboards with one command — opencode auth login auto-configures models, MCP servers, agents and permissions from a discovery endpoint; a wrangler deploy updates 3,000+ people, reaching 93% R&D adoption in under a year (first-party) — blog.cloudflare.com
Expedia Group runs hundreds of MCP servers, tools and skills in one searchable, graded catalog with SSO, per-asset security scanning and SOC 2-grade audit; engineers find vetted capabilities in minutes (first-party, AWS blog) — aws.amazon.com
The open SKILL.md standard and the .agents/skills/ convention make skills installed by one compliant client automatically visible to others — the portability layer for any multi-harness future (official spec) — agentskills.io
Cisco reportedly pre-installs org-scoped plugins via policy states (INSTALLED_BY_DEFAULT / AVAILABLE / NOT_AVAILABLE) for 2,000+ engineers, with SIEM-streamed audit — a secondary-source claim, not independently verified — baeseokjae.github.io
Tencent's TeamAI keeps everything as Git content in .teamai/: three-command onboarding, per-agent adapters writing into each harness's expected paths, PR review and pre-push secret scanning (secondary source) — topdigg.com
Your current npx approach: keep, change, add
Keep
One command: npx @company/agent-skills stays the whole happy path — Cloudflare's entire onboarding is one auth command, Tencent's is three
Artifactory and Git as the trusted supply chain
Central repository as the single source of packages
Change
Installer stops copying files ad hoc; it authenticates, registers the internal catalog, picks or infers a role profile, writes a lockfile
Skills move to the portable .agents/skills/ SKILL.md convention; OpenCode-specific agents, commands and hooks go behind an adapter
Updates become atomic and automatic (startup or background sync) instead of re-running the installer
Add
Signed, versioned catalog with owner, semver, compatibility range, permissions, changelog, review date, deprecation status per package
Stable and beta channels, one-command rollback, in-agent search/install/update/doctor
Remote MCP gateway with SSO for shared tools — no server processes or credentials on laptops
Adoption and failure telemetry (activation, active use, stale packages), not prompt contents
Roadmap in three phases
Phase 1 — formalize
2–4 weeks
Versioned manifests and ownership for every package
SKILL.md where portable
CI schema, lint and evaluation tests; secret scanning
Lockfile; preserve the existing command unchanged
Phase 2 — control plane
4–8 weeks
Signed catalog metadata and role/team profiles
Startup sync, stable/beta channels, rollback
In-agent search / install / update / doctor
Phase 3 — operate
ongoing
Shared tools become remote MCP behind SSO gateway
Approvals, audit logs, kill switches, telemetry
Additional harness adapters without duplicating content
Checklists and success criteria
Every package manifest carries
Owner and CODEOWNERS-backed review
Semantic version and compatibility range
Permissions and capabilities declared
Changelog, review date, deprecation status
Signature or checksum; ideally SBOM/provenance
Governance baseline
Pull-request review; CI validation and evaluation tests
Secret and supply-chain scanning before publish
Deny-by-default for sensitive MCP tools; least privilege via SSO groups
Centralized secrets — none installed on laptops
Kill switches and audit logs
Measure success as
Clean machine to first successful use: under 5 minutes (median)
At least 80% of target developers on current stable profile
Update success above 99%; rollback under 2 minutes
Package ownership and review coverage: 100%
Declining support tickets and duplicated skills
Developer enablement as a product
Approved default profile installs automatically
Searchable in-agent catalog for opt-in extras
Documentation inside the agent, curated starter profiles
Office hours and a feedback path
All 20 evidence rows: platforms and company implementations
Platform / company
Evidence
Centralization model
Install & discovery
Governance / result
Source
Method: 20 researched rows (10 native platform mechanisms, 10 company implementations), one row per selected platform or company, distilled from 1,544 official-document and 1,910 company-practice candidates as of 2026-09-24. Columns describe each source's centralization model, install/discovery flow, governance and reported assets or results; long cells are truncated in the table with full text on hover. Badges mark official product documentation, first-party engineering sources, and secondary-source claims (e.g. Cisco metrics), which are not independently verified. Architecture diagram is the recommended synthesis, not a single source's design.