The MCP and AI-tool-integration ecosystem

1) Thesis

MCP is restructuring AI-tool integration from bespoke, point-to-point agent connectors into a standardized, policy-controlled “tool plane,” shifting value toward trusted tool distribution, identity/security, observability, and ownership of proprietary enterprise workflows.

2) Market map

Segment 1: Protocols, clients, and agent-development frameworks

Companies: Anthropic, OpenAI, Microsoft, Cursor, LangChain, LlamaIndex

Dynamics: MCP began as Anthropic’s open protocol and is being absorbed into major model, IDE, and agent frameworks; protocol access is likely to commoditize, while distribution through developer environments and agent runtimes remains strategically valuable.

• Anthropic — creator of MCP and its reference ecosystem.
• OpenAI — supports MCP-style tool connectivity in its agent/API stack. [MCP support: high confidence]
• Microsoft — MCP support across parts of its Copilot, Azure, and developer ecosystem. [High confidence, although product coverage is evolving]
• Cursor — developer IDE with MCP integration.
• LangChain — agent framework with MCP-related integrations.
• LlamaIndex — agent/data framework with MCP support. [High confidence for ecosystem integration; exact product scope is evolving]

---

Segment 2: Tool servers, connectors, and MCP marketplaces

Companies: Zapier, Composio, Pipedream, Apify, Smithery, Mintlify

Dynamics: These companies convert APIs, SaaS actions, data sources, and web workflows into agent-consumable tools; the market is relatively easy to enter, so differentiation will come from breadth, reliability, permissions, and distribution rather than connector count alone.

• Zapier — exposes automation actions and workflows to AI agents, including MCP-oriented tooling.
• Composio — developer platform for authenticated tools and integrations for agents.
• Pipedream — API and workflow infrastructure with agent/MCP integrations.
• Apify — web-scraping and automation actors that can be exposed as agent tools.
• Smithery — MCP server directory/registry and discovery layer.
• Mintlify — documentation infrastructure and MCP-related tooling for developers. [Less certain as a major standalone MCP platform]

---

Segment 3: MCP gateways, identity, security, and enterprise control planes

Companies: Cloudflare, Kong, Gravitee, Solo.io, Auth0, Lasso Security

Dynamics: This is becoming the enterprise choke point: organizations need centralized authentication, authorization, rate limits, audit logs, data-loss prevention, tool approval, and runtime isolation before allowing agents to access production systems.

• Cloudflare — MCP server and gateway capabilities, with strong edge, identity, and security distribution.
• Kong — API gateway and AI gateway products with MCP-related support.
• Gravitee — API management and AI/MCP gateway capabilities.
• Solo.io — service-mesh/API infrastructure with MCP gateway positioning. [MCP product scope is evolving]
• Auth0 — identity and authorization infrastructure relevant to securing MCP servers and agent tools. [Adjacent rather than a pure-play MCP vendor]
• Lasso Security — AI security tooling focused on agent and MCP risks. [Less certain on scale and breadth relative to the larger gateway vendors]

---

Segment 4: Agent runtimes, orchestration, and model-serving platforms

Companies: AWS, Microsoft, Google, LangChain, LlamaIndex, CrewAI

Dynamics: Agent runtimes determine how tools are selected, invoked, retried, composed, and evaluated; hyperscalers have distribution advantages, but independent frameworks remain important where customers want model and infrastructure portability.

• AWS — Amazon Bedrock Agents and related agent/tool orchestration capabilities.
• Microsoft — Copilot Studio, Azure AI Foundry, and agent development infrastructure.
• Google — Vertex AI Agent Builder and Google Agent Development Kit.
• LangChain — widely used open-source agent orchestration layer.
• LlamaIndex — data-connected agent and retrieval framework.
• CrewAI — multi-agent orchestration framework with tool integrations. [MCP support and commercial traction are evolving]

---

Segment 5: Agent observability, evaluation, and runtime governance

Companies: Langfuse, Arize AI, Braintrust, Helicone, Datadog, Weights & Biases

Dynamics: MCP increases the number of external actions an agent can take, making tracing, cost attribution, tool-level evaluation, policy enforcement, and incident investigation more important; observability is likely to consolidate into broader developer and security platforms.

• Langfuse — open-source LLM and agent tracing/evaluation platform.
• Arize AI — AI observability and evaluation through Phoenix and enterprise products.
• Braintrust — LLM evaluation, testing, and tracing.
• Helicone — LLM gateway, logging, and cost/usage analytics.
• Datadog — expanding into LLM and AI-agent observability. [MCP-specific coverage is less certain]
• Weights & Biases — model and experiment management with growing LLM/agent evaluation capabilities. [Adjacent rather than MCP-specific]

3) Segment likely to capture the most value

Winner: Enterprise control planes for identity, security, governance, and observability

The gateway/control-plane layer is most likely to capture disproportionate value because it sits directly on the enterprise approval path and can become mandatory infrastructure rather than an optional developer convenience. It can monetize:

• Authentication and delegated authorization
• Tool approval and allowlisting
• Secrets management
• Tenant isolation and sandboxing
• Audit trails and compliance
• Rate limiting and cost controls
• DLP and prompt/tool-injection defenses
• Reliability, tracing, and incident response

The protocol itself should be broadly available and low-margin. Connector catalogs will face commoditization. Agent frameworks have strong developer influence but are vulnerable to hyperscaler bundling. By contrast, the control plane can become the policy and traffic layer through which every agent-to-tool interaction must pass.

A caveat: the ultimate economic value remains with companies that own proprietary data and business workflows—such as Salesforce, ServiceNow, SAP, Workday, Stripe, and major vertical software providers. The control plane is the most attractive new infrastructure segment; systems of record retain much of the underlying application value.

4) White-space opportunities

1. Neutral, enterprise-grade MCP trust and permission infrastructure

There is no clearly dominant, vendor-neutral layer that lets enterprises:

• Discover and inventory MCP servers
• Verify server provenance and software supply-chain integrity
• Assess tool risk before approval
• Bind tools to user, agent, tenant, and data permissions
• Enforce fine-grained, purpose-limited authorization
• Revoke access instantly
• Produce compliance-grade records of every tool call

Existing API gateways, IAM products, and AI-security tools each cover portions of this problem, but no obvious player owns the complete MCP server registry + trust scoring + delegated authorization + runtime enforcement workflow.

2. Durable, transactional agent-action infrastructure

Most MCP tooling focuses on exposing functions, not safely executing consequential business actions. A missing layer would provide:

• Idempotency and duplicate-action prevention
• Human approval checkpoints
• Transaction previews and simulations
• Rollback or compensating actions
• Long-running job handling
• Versioned tool contracts
• SLA and failure semantics
• Financial and operational reconciliation

This would be especially valuable for payments, procurement, customer support, healthcare, and enterprise operations. Existing workflow and automation platforms address parts of it, but there is no clear dominant product designed specifically for auditable, reversible, multi-step agent transactions across MCP tools.