AI Agent Marketplaces and Distribution Channels: The Race to Own Agent Commerce

Executive Summary

The distribution of AI agents is fragmenting across five distinct marketplace models, each targeting different segments of a market projected to reach $52 billion by 2030. Anthropic's Claude Marketplace (launched March 7, 2026) takes a zero-commission enterprise procurement approach, while OpenAI's GPT Store — with 3 million GPTs created — delivers disappointing ~$0.03/conversation payouts to most creators. Salesforce's AgentExchange leads in enterprise traction with $800M ARR and 18,500 customers. Meanwhile, open protocols (MCP, A2A, Agent Skills) are converging under the Agentic AI Foundation to prevent platform lock-in. The central tension: centralized marketplaces offer trust and discoverability, while open protocols promise interoperability and freedom. Security remains the critical gap — 88% of organizations report AI agent security incidents, and 90% of deployed agents are over-permissioned.

The Five Marketplace Models

Agent distribution in 2026 follows five distinct patterns, each optimized for different buyer journeys and value chains.

Claude Marketplace: Enterprise Procurement Layer

Anthropic's Claude Marketplace, launched just days ago on March 7, 2026, takes the most unconventional approach. Rather than building a consumer app store, Anthropic modeled it after AWS Marketplace — enterprises with existing spend commitments can redirect budget toward third-party tools built on Claude.

The headline differentiator: zero percent commission. While AWS charges 3-15% and Azure takes similar cuts, Anthropic forgoes transaction revenue entirely. The strategy is clear — deepen enterprise lock-in through ecosystem breadth rather than extract per-transaction value. Six launch partners (Snowflake, GitLab, Harvey AI, Rogo, Replit, Lovable Labs) span legal automation, financial analysis, developer tools, and no-code building.

This is not a browseable store. Discovery happens through account teams, not search bars — a deliberate choice that trades consumer-style virality for enterprise procurement alignment.

GPT Store: The Creator Economy Meets Reality

OpenAI's GPT Store, launched in January 2024, has achieved impressive scale — 3 million GPTs created, 159,000 publicly active — but monetization tells a sobering story. The engagement-based revenue share works out to approximately $0.03 per conversation. At that rate, earning $1,000/month requires 33,000+ quality conversations, with a minimum threshold of 25 conversations per week just to qualify.

The real money, it turns out, is not in the store. B2B consulting — building custom enterprise GPTs for $5,000-$20,000 setup fees plus monthly maintenance — generates far more revenue than marketplace distribution.

OpenAI is building a second layer to address this: the ChatGPT App Directory, accepting developer submissions since December 2025. Built on MCP (Anthropic's protocol, notably), the Apps SDK enables richer integration — external service connections, real-time context fetching, action triggering, and in-chat UI rendering. This two-tier model (casual GPTs + serious apps) may become the industry template.

With 700 million weekly ChatGPT users and 81% market share, OpenAI's distribution advantage is formidable. Analysts estimate the ChatGPT App Directory could redirect $44 billion annually from traditional app stores. But discoverability in a sea of 159,000 GPTs remains an unsolved problem.

Microsoft Agent Store: IT-Gated Enterprise Distribution

Microsoft's Agent Store, integrated within Microsoft 365 Copilot, launched with 70+ agents spanning Teams, Outlook, Word, Excel, and PowerPoint. The distribution model is explicitly IT-admin-gated — developers publish to organizational catalogs or the Microsoft Commercial Marketplace, then IT administrators enable agents for their organizations.

Two development paths cater to different builder profiles: Copilot Studio for low-code/no-code builders using natural language, and the Microsoft 365 Agents Toolkit for professional developers using the Agents SDK with Visual Studio Code integration.

The strategic advantage is obvious: agents surface where enterprise workers already live. Rather than asking users to visit a separate store, agents appear contextually across the M365 surface area.

Google Agentspace: The Agent Finder

Google's approach combines a Cloud Marketplace listing (traditional procurement) with Agentspace, a discovery platform featuring an AI-powered Agent Finder that helps employees discover the right agent for their task. The "Shopify for agents" tagline signals ambitions beyond simple listing — Google wants to own the agent discovery experience itself.

Partners include Accenture, Deloitte, PwC, and EPAM. PwC is building an "expansive AI agent ecosystem" with Google Cloud, while EPAM has launched 7 advanced agents on the marketplace. Google plans to add hundreds of additional agents in coming months.

Salesforce AgentExchange: The Revenue Leader

Salesforce's AgentExchange, launched at TDX '25 in March 2025, has the strongest adoption numbers of any agent marketplace. By Q4 FY2026: 18,500 customers (9,500+ on paid plans), 29,000+ cumulative deals rising 50% quarter-over-quarter, and ARR reaching $800 million — up 169% year-over-year. It is the fastest-growing organic product in Salesforce history.

The pricing evolution is instructive. Salesforce started with a simple $2/conversation model, but complex multi-step workflows made this confusing for buyers. By 2026, they've evolved to three tiers: Conversations ($2/conversation), Flex Credits (100,000 credits for $500 with per-action metering), and Per-User ($125/user/month for unlimited employee-facing usage). This progression from simple to multi-tier mirrors the market's search for the right pricing paradigm.

With 200+ initial partners (including Google Cloud, DocuSign, Box) and 1,000+ trained to build agents, AgentExchange has the deepest ecosystem bench.

Protocol-Based Discovery: The Open Alternative

Beyond centralized marketplaces, a parallel distribution model is emerging through open protocols — agents discovering each other programmatically rather than through human-curated stores.

The Converging Protocol Stack

Five protocols operate at different layers, and consolidation is already happening:

MCP (Model Context Protocol) handles the vertical agent-to-tool connection. Originally Anthropic's, now donated to the Agentic AI Foundation (AAIF) under the Linux Foundation. With 1,000+ community-built servers and adoption by OpenAI, Google, and Microsoft, MCP is the de facto standard for tool access. Its critics note that authentication, authorization, and sandboxing are "left to the implementer — most skip all three."

A2A (Agent-to-Agent Protocol) handles horizontal inter-agent communication. Google-created with 100+ supporting enterprises by February 2026. IBM's competing ACP (Agent Communication Protocol) has already merged into A2A, signaling consolidation at this layer. Each A2A server publishes a JSON Agent Card at /.well-known/agent.json — a machine-readable capability description enabling programmatic discovery.

Agent Skills handles cross-platform portability. Anthropic's December 2025 specification, now also under AAIF, enables a skill built for Claude to run on ChatGPT or Copilot. Already adopted by Microsoft, OpenAI, Atlassian, Figma, Cursor, and GitHub. This commoditizes the "skills" layer, ending the era of proprietary GPTs and Actions.

The practical path for implementers: start with MCP for tool access, add A2A when you need multi-agent collaboration, and use Agent Skills for cross-platform distribution.

The Agentic AI Foundation

Formed in December 2025 under the Linux Foundation, AAIF provides neutral governance to prevent protocol fragmentation. The founder list reads like a détente treaty: Anthropic, OpenAI, and Block as founders, with Google, Microsoft, AWS, Bloomberg, and Cloudflare as supporters. Inaugural projects include MCP, Block's open-source agent framework Goose, and AGENTS.md.

This is remarkable cooperation in a fiercely competitive market. The lesson: standards benefit all players when the alternative is ecosystem fragmentation and enterprise buyer paralysis.

The Security Crisis

The speed of agent deployment has outpaced security infrastructure, creating what may be the most significant near-term risk to the entire marketplace ecosystem.

The numbers are alarming: 88% of organizations report confirmed or suspected AI agent security incidents. 90% of deployed agents are over-permissioned, holding 10x more privileges than required. Only 14.4% of organizations report all AI agents going live with full security and IT approval — yet 80.9% of technical teams have moved past planning into active testing and production.

Shadow AI compounds the problem: the average enterprise has approximately 1,200 unofficial AI applications in use, with 86% reporting no visibility into AI data flows. Shadow AI breaches cost $670,000 more on average than standard security incidents.

Package registries face their own challenges. Researchers discovered 1,184 malicious skills in one major registry — approximately 1 in 5 packages — alongside 135,000 instances exposed to the public internet with insecure defaults.

Regulatory response is accelerating. NIST launched its AI Agent Standards Initiative on February 17, 2026, with three pillars: industry-led standards development, community-led open source protocol development, and AI agent security and identity research. The EU AI Act reaches full enforcement for high-risk systems in August 2026. OWASP has published its AI Agent Security Top 10 for 2026, with prompt injection as the number one risk.

Marketplaces that solve trust will win. Expect "verified agent" programs, security certifications, and compliance badges to become table-stakes differentiators.

Enterprise vs. Consumer: Two Different Worlds

Enterprise Distribution

Enterprise agent distribution is converging on three requirements: compliance-first deployment, scoped access control, and audit trails that capture not just the "what" (actions taken) but the "why" (prompt, context, decision logic).

Organizations with evidence-quality audit trails score 20-32 points higher on every AI maturity metric. GitHub's Enterprise AI Controls, now generally available, includes an agent control plane with audit log filtering by specific agents — including third-party agents.

The governance-containment gap is the central challenge: most organizations can monitor agents but cannot stop them when something goes wrong. Agent identity management — treating agents as first-class IAM entities with scoped tokens limited to exact functions — is becoming a prerequisite for enterprise deployment.

Consumer Distribution

Consumer agent distribution faces a different problem: attention without retention. OpenAI has released multiple new experiences (Pulse, Group Chats, Record), but none have truly "broken through" in sustained usage. Several waves of AI app virality in 2025-2026 produced moments of attention without sticky daily habits.

Market concentration adds friction: fewer than 10% of ChatGPT weekly users visit another model provider. Consumer trust remains fragile — only about one-third are willing to complete payment through an AI answer engine, with data privacy as the primary concern.

Consumer agent distribution may follow mobile app dynamics: a few breakout hits, a long tail of obscurity, and discovery power concentrating in the platform that owns the default interface.

Market Outlook

The AI agents market is estimated at $7.8-10.9 billion in 2026, projected to reach $52 billion by 2030 at approximately 43% CAGR. Gartner predicts 40% of enterprise applications will include task-specific AI agents by 2026, and 15% of work decisions will be made autonomously by AI agents by 2028.

But scaling remains the central challenge. Gartner also predicts more than 40% of agentic AI projects will be canceled by end of 2027 — the gap between experimentation (~65% of organizations) and scaled production (<25%) is where most projects die.

The competitive landscape is bifurcating: enterprise distribution is being won by cloud procurement layers (Anthropic, AWS, Azure, GCP) and platform-native marketplaces (Salesforce), while consumer distribution remains OpenAI's game with 81% market share. Open protocols (MCP, A2A, Agent Skills) serve as connective tissue, enabling agents built for one platform to operate on another — if the standards governance holds.

The open question for 2027: Will there be a dominant "agent app store" like Apple's App Store, or will distribution remain fragmented across cloud providers, enterprise platforms, and open protocols? The answer likely depends on whether security and trust can be standardized — because in a market where 88% of organizations report security incidents, the marketplace that solves trust owns the distribution.