MCP Integration Platforms

Related: For platforms that deploy the agents themselves, see Team Agent Platforms Compared. For agent development frameworks, see Agent Frameworks Compared.

Executive Summary

MCP integration platforms solve the "last mile" problem for AI agents: authenticated access to the SaaS tools where real work happens. Rather than every team building OAuth flows and API wrappers, these platforms provide hosted MCP servers, pre-built tool catalogs, and managed credential lifecycles — so agents can act in GitHub, Slack, Salesforce, and thousands of other apps without credentials ever entering the model's context.

Key Findings:

• The category got its validation event — Workday's acquisition of Pipedream (announced November 2025) signals that agent tool access is enterprise infrastructure, not developer tooling^[1]
• Auth model is the real differentiator — Zapier MCP acts as the account owner; Pipedream, Arcade, and Nango manage per-end-user OAuth; Paragon manages per-tenant auth for embedded agents^[2][3]
• Context-window economics matter — exposing thousands of tools bloats agent context; Klavis's Strata progressively discloses tools in stages and benchmarks above official single-app servers^[4]
• The spec itself is contested ground — Arcade co-authored the URL Elicitation proposal with Anthropic, standardizing in-conversation OAuth for MCP servers^[5]

Strategic Planning Assumptions:

• By 2027, per-end-user authorization will be a procurement requirement for enterprise agent deployments
• By 2027, at least one more MCP integration platform will be acquired by an enterprise software vendor following Workday's playbook
• By 2028, first-party directories will serve the top 100 integrations, leaving independent platforms to compete on auth, governance, and the long tail

---

Market Definition

MCP integration platforms give AI agents authenticated access to third-party SaaS tools. Core capabilities:

• Hosted MCP servers — tools exposed over the Model Context Protocol to any compliant client
• Managed auth — OAuth flows, token refresh, and credential storage handled by the platform; credentials never reach the model
• Tool catalogs — pre-built, maintained integrations rather than hand-rolled API wrappers
• Scoping — sessions scoped to an end user, tenant, or account depending on the platform

Inclusion Criteria:

• Primarily an integration/tool-access layer for AI agents (not an agent framework or client)
• Ships managed MCP servers and/or agent-native tool APIs with managed auth
• Real shipped product with meaningful traction (funding, stars, or vendor backing)

Exclusion Criteria:

• Agent frameworks (LangChain, CrewAI) — consume these platforms rather than compete with them
• Single-app MCP servers and registries/directories (Smithery, Glama, the official MCP registry) — discovery and distribution, not managed access
• Enterprise MCP gateways (TrueFoundry, Portkey, Turbo MCP) — govern servers others build; adjacent category
• First-party directories (Anthropic's connectors, Google's managed MCP servers^[6]) — distribution channels from model/cloud vendors, not independent platforms

---

Comparison Matrix

---

How the Market Segments

Adjacent but excluded: registries/marketplaces (Smithery's ~7,000 servers, Glama's directory), enterprise MCP gateways (TrueFoundry, Portkey), and first-party channels (Anthropic's connectors directory, Google's 50+ managed MCP servers^[6]). Watchlist for future inclusion: ACI.dev, StackOne, Merge's Agent Handler, Smithery.

---

Platform Profiles

Arcade

Auth-first MCP runtime — agents act as the authenticated end user^[7]

• User-scoped OAuth + IdP integration — agents get the end user's permissions, not a service account's
• Spec influence — co-authored URL Elicitation with Anthropic, now in the MCP spec^[5]
• 8,000+ prebuilt tools plus an open-source MCP server framework with evals built in
• Seed stage — $12M (Laude Ventures); customers include LangChain and Snyk

Best For: Enterprises that need agents acting under individual user permissions with IdP integration.

---

Composio

Open source agent integration layer — 1000+ toolkits, managed auth, 27K stars^[8]

• 1000+ toolkits built agent-native, with managed OAuth and sandboxed execution
• Framework-agnostic — OpenAI, Anthropic, LangChain, CrewAI, AutoGen, Vercel AI SDK
• Skill learning — tool-usage knowledge improves across agents^[9]
• Best-funded independent — $29M (Lightspeed-led Series A, July 2025)

Best For: Teams building custom agents that need broad SaaS coverage with agent-native ergonomics.

---

Klavis AI

Open-source MCP platform with context-aware tool exposure — Strata unified server^[10]

• Strata progressive discovery — exposes servers → categories → actions → schemas in stages to fit the context window; outperforms official GitHub/Notion servers on MCPMark^[4]
• 300+ services with managed OAuth; SOC 2 Type II
• MCP sandboxes for agent training and RL
• Early — YC X25, team of 3; the most direct open-source Composio competitor

Best For: Teams hitting context-window limits with large tool catalogs; open-source-first builders.

---

Nango

Auth-first integration infrastructure with a hosted MCP server — 800+ APIs^[11]

• Managed OAuth across 800+ APIs — connections, token refresh, webhooks, authenticated proxy
• Per-connection MCP scoping — each MCP session bound to one user connection; credentials never enter the agent runtime
• Code-first and self-hostable (Elastic License 2.0); 10K+ engineering teams, customers include Replit and Ramp
• Capital-efficient — $7.5M seed (Gradient, April 2026), cash-flow positive pre-raise^[12]

Best For: Engineering teams that want auth infrastructure under their control, with MCP as one consumption surface.

---

Paragon

Embedded iPaaS rebuilt for AI — ActionKit gives your product's agents 1,000+ actions^[3]

• Per-tenant managed auth — each of your customers connects their own SaaS accounts
• ActionKit as API or MCP server — 130+ integrations, 1,000+ actions, listed in Anthropic's registry
• Built for B2B SaaS — agents inside your product, not your internal tooling
• Quote-based pricing — metered on connected users; founded 2019, SOC 2 Type II

Best For: B2B SaaS companies shipping agents inside their own product that act on customers' connected tools.

---

Pipedream

Per-end-user MCP servers over a 3,000-app catalog — now part of Workday^[13]

• Dedicated MCP servers per app, per end user via Connect; fully managed end-user OAuth
• 10,000+ pre-built tools on a workflow platform with 5,000+ customers
• Workday acquisition (announced Nov 2025) — category validation, but roadmap uncertainty for third-party developers^[1]
• Credit-based pricing; Connect free in dev mode

Best For: Teams wanting the deepest per-end-user MCP catalog — with eyes open about Workday's roadmap priorities.

---

Zapier MCP

The incumbent's entry — 9,000+ apps, 30,000+ actions through one endpoint^[2]

• Largest catalog in the category by an order of magnitude
• Hosted MCP server per user with first-party Claude and ChatGPT setup flows
• Account-owner credentials — no per-end-user OAuth; weaker for embedded/multi-user scenarios
• Zero vendor risk — profitable parent; but agent infrastructure is not Zapier's core business

Best For: Individuals and teams wanting maximum app coverage for agents acting under one account.

---

Gap Analysis

Key gaps across the market:

• No platform combines Zapier-scale catalog breadth with per-end-user auth — Pipedream comes closest
• Governance is immature — audit trails, RBAC, and approval flows for agent tool calls are early everywhere; enterprise gateways are filling the gap from the outside
• Pricing is hard to compare — connections, tool calls, tasks, credits, and connected users all meter differently; expect normalization pressure

---

Strategic Recommendations

By Use Case

By Buyer Profile

Solo builders / prototypes: → Zapier MCP for instant coverage; Klavis AI's free tier for open-source work.

Startups building agent products: → Composio or Klavis AI for tool breadth; Nango if integrations are core to your product and you want code-level control.

B2B SaaS adding agents to your product: → Paragon's ActionKit for per-tenant auth; Nango for a code-first alternative.

Enterprises deploying internal agents: → Arcade for IdP-integrated, user-scoped authorization; evaluate Pipedream with Workday's roadmap in mind.

---

Market Outlook

Near-Term (2026)

• Workday closes the Pipedream acquisition; expect enterprise-suite bundling and at least one competitive response acquisition
• Per-end-user authorization patterns (URL Elicitation) ship in mainstream MCP clients
• First-party directories keep absorbing the most popular integrations^[6]

Medium-Term (2027-2028)

• Auth-first and catalog-first platforms converge — auth players add catalogs, catalog players deepen auth
• Agent tool-call governance (who authorized what, when) becomes a compliance requirement, favoring platforms with audit-grade logging
• Registry/marketplace players consolidate or get absorbed by platforms and model vendors

Long-Term (2029+)

• Managed tool access becomes default infrastructure, like payment processing — built by few, used by all
• The MCP spec (now under the Linux Foundation) becomes the assumed substrate; differentiation moves entirely to auth, governance, and reliability

---

Bottom Line

MCP integration platforms are where agent ambition meets SaaS reality. The category's defining question is whose credentials does the agent hold? Account-owner models (Zapier MCP) are simplest; per-end-user models (Arcade, Pipedream, Nango) are what enterprises will require; per-tenant models (Paragon) serve products embedding agents.

Current Leaders:

• Composio for agent-native tool breadth with open source and the strongest independent funding
• Pipedream for per-end-user MCP depth — now with Workday's resources and Workday's priorities
• Zapier MCP for raw coverage under a single account

Rising Challengers:

• Arcade for enterprise-grade, IdP-integrated authorization and direct influence on the MCP spec
• Klavis AI for context-efficient tool exposure (Strata) as catalogs outgrow context windows
• Nango for capital-efficient, code-first auth infrastructure

Start with the auth model your deployment requires, then choose the largest catalog that supports it.

---

Research by Ry Walker Research • methodology