Skip to main content
Tutorial4 min read

Internal agent marketplace setup: ship your company's agents like apps

Internal agent marketplaces let employees discover, install, and use approved agents like an app store. Architecture, governance, distribution, and the playbook for the first 90 days.

Internal "agent marketplaces" are showing up in large enterprises in 2026 — the App Store metaphor applied to AI agents inside the company. Employees browse, install, and use approved agents on their own. Done well, it ten-x's adoption while tightening control. Here is the build.

Why an internal marketplace beats per-team rollout

Three problems it solves:

  • Discovery — most employees do not know what agents already exist.
  • Quality — every team building solo means duplicated work and inconsistent quality.
  • Governance — central listing means central review.

The marketplace becomes the choke point and the catalogue at once.

The components

Five layers:

1. Catalogue

A web app where employees browse agents. Per agent: description, owner, supported tasks, rating, last updated.

2. Submission and review

A pipeline for teams to submit new agents. Review by the AI Risk Committee (see governance) before listing.

3. Distribution

A one-click install: adds the agent to the user's available tools in Claude Desktop, Cursor, or your in-house agent host.

4. Telemetry

Per-agent usage, satisfaction, error rates. Feeds the observability platform.

5. Lifecycle management

Versioning, deprecation, retirement.

Build vs buy

Three patterns in 2026:

  • Build on top of the self-hosted MCP registry — most flexible, ~3 engineer-months.
  • Bolt on a marketplace product — Workday, ServiceNow, Atlassian all shipping AI marketplace features.
  • Adopt a vendor — Anthropic Enterprise, Microsoft Copilot Studio, smaller niche players.

Most large enterprises start by building because none of the products quite fit their identity stack and governance.

The first 90 days

A working roll-out:

  • Week 1–2: define charter, governance hooks, telemetry shape.
  • Week 3–4: pilot with three internal agents (one per business unit).
  • Week 5–8: open submission to all teams. Tight review.
  • Week 9–12: broaden access; introduce ratings; first deprecations.

After 90 days you have signal on demand patterns, common requests, and governance bottlenecks.

Governance hooks

Three integrations to build before launch:

  • Risk-tier gating — the marketplace shows only agents the user is approved to use.
  • Audit logging — every install, use, and uninstall feeds the audit trail.
  • Kill switch — operator can pull an agent from the catalogue and revoke active installs.

Identity and access

Single-sign-on into the marketplace; per-user entitlements based on group membership. See agent SSO patterns.

The marketplace itself does not see user data. It surfaces metadata and routes to per-agent backends.

Pricing and cost attribution

Even internally, cost matters. Three approaches:

  • Free use — central IT pays. Simple. Encourages adoption. Loses cost visibility.
  • Show-back — usage is metered per BU; no transfer; visibility only.
  • Charge-back — usage transfers cost to the BU.

Most enterprises start free, move to show-back at six months, charge-back at twelve.

Quality signals in the catalogue

Three signals per agent listing:

  • Stars (1–5 from users).
  • Active install count.
  • SLO status (green / amber / red from monitoring).

Agents with red SLO get demoted automatically. Stars below threshold triggers a steward review.

Lifecycle: deprecation done right

Three phases:

  1. Mark deprecated — visible warning; new installs paused.
  2. Sunset notice — 90-day countdown for existing users.
  3. Remove — uninstall script removes from active hosts.

Skipping the middle phase produces user revolts.

Common mistakes

  • No review on submission — marketplace fills with low-quality forks.
  • No sunset process — agents accumulate; the catalogue becomes noise.
  • No usage signals — operators cannot tell which agents matter.
  • Underestimating identity work — entitlements take longer than the catalogue itself.

What to watch for

A successful marketplace produces:

  • 70%+ of the org adopting at least one agent within 6 months.
  • 3–5x reduction in shadow-AI tools (employees stop installing their own).
  • Clear top-10 most-used agents that become company standards.

If you do not see these by month 6, the discovery or distribution layer needs rework.

Where this is heading

Three trends by 2027: turnkey enterprise agent marketplace products from major SaaS vendors, MCP-native marketplace standards, and cross-org marketplaces (industry consortia sharing approved agents). Build with portable abstractions now.

Loadout

Build your AI agent loadout

The directory of MCP servers and AI agents that actually work. Pick the right loadout for Slack, Postgres, GitHub, Figma and 20+ integrations — with install commands ready to paste into Claude Desktop, Cursor or your own stack.

© 2026 Loadout. Built on Angular 21 SSR.