MCP Integration Platforms

Related: Agent frameworks build the agent loop; team agent platforms operate agent work. This report covers the reusable tool-access layer between an agent and external systems.

Executive Summary

An agent that can name an API operation still needs a valid account, appropriate permission, a usable input schema, and evidence that the operation succeeded. MCP integration platforms package parts of that work: maintained connectors, authorization flows, tool discovery, execution, and operational controls. This comparison evaluates ten platforms, with primary sources checked September 16, 2026.

The most consequential updates change the old category boundaries. Zapier now documents embedded end-user MCP authentication through White Label. Smithery supplies managed connections, not just registry discovery. Klavis's current site emphasizes agent-training data while its Strata integration documentation remains available. Workday has confirmed that its Pipedream acquisition closed, and Arcade announced its acquisition of Smithery in August.^{[1][2][3][4][5]}

For readers trying to complete engineering work, Tembo is relevant at the execution layer: coding agents use repositories and connected tools to produce reviewable work. Its role is explained below rather than treating it as an interchangeable connector SDK.^[6] Disclosure: Ry Walker is Tembo's CEO and co-founder.

---

Market Definition

A member must provide reusable access to multiple third-party applications through managed MCP or agent-oriented tool APIs, with documented authentication and an implementation surface a team can evaluate. Funding, stars, and catalog totals are not entry requirements. This revision retains the original seven members and adds StackOne, Merge Agent Handler, and Smithery, each with an individual profile.

A single-app server is narrower. A registry that only indexes servers does not meet the access requirement. A general agent framework or engineering execution platform solves a different part of the workflow. Products can overlap: Smithery qualifies because of Connect, while a governance product qualifies only when it also supplies the integration capability being compared.

Klavis remains a qualified legacy offering in the matrix because Strata's hosted and local connection paths are still documented. Its changed public focus warrants a support and commercial check, not an unsupported declaration that the product has shut down.^[7][3]

Adjacent: Grok as an MCP Client

SpaceXAI / Grok API consumes remote MCP servers: the caller supplies a Streaming HTTP or SSE URL, optional authorization/headers, and permitted tool names. An omitted or empty allowlist makes all of the server's tools available. Its documented Responses compatibility excludes require_approval and connector_id.^[8]

This is relevant connectivity, but it does not by itself supply the maintained connectors and customer authorization lifecycle required for membership here. Grok can be the client of an integration service; that does not make the two interchangeable. The ten-member matrix is unchanged by this September 16 source check.

For a proposed ticket-triage workflow, first establish whose account the remote server uses, which tools it exposes, and where a write is authorized. A tool-name allowlist narrows the model's choices; it does not establish a per-action confirmation step. Treat this as application design guidance: implement approval in a controlled execution path when required, instead of assuming an unsupported parameter will enforce it. The framework, integration provider, and an engineering execution platform such as Tembo each need an explicit role in that design.^[6]

Comparison Matrix

The final column identifies a concrete evaluation issue. It is not an assertion that the other vendors lack that capability.

---

Platform Profiles

Arcade

Arcade is a useful candidate when the organization needs a tool runtime with user verification, governance hooks, and commercial private-deployment options. Its contextual-access design supports separate hooks for tool visibility, request checks, and output transformation. The customer supplies the relevant policy logic.^[20]

A material implementation requirement is easy to miss: Arcade's default OAuth apps pair with its own user verifier and project membership. Its production guide directs multi-user applications to use their own OAuth credentials with a custom verifier.^[9] Budget for that application integration rather than extrapolating from a successful developer login.

Best fit: Teams that need explicit authorization and organization policy around custom and packaged tools. Confirm the required deployment and support tier; enterprise terms include private and air-gapped options.^[21]

Composio

Composio organizes a run around a user session that combines available toolkits, connected accounts, authentication, and execution state. Meta tools discover and execute relevant actions instead of loading the whole catalog into the model context. Toolkits are discoverable broadly by default, so a narrower application should configure its allowed set.^[22]

Its B2B guide distinguishes a member's account from a workspace-owned integration. Collapsing every employee into one tenant ID would share the same connection scope; an application must enforce its own member and workspace authorization. Explicit shared connections are documented as experimental.^[10]

Best fit: Custom agents needing packaged SaaS access and reusable task context. Check whether the intended route uses a managed OAuth app, the customer's own app, or directly supplied tokens, because both responsibility and billing differ.^[23]

Klavis AI

Klavis AI needs a status-qualified recommendation. Its current homepage promotes coding and tool-use data for training agents. Its Strata quickstart still describes a managed endpoint that selects services, associates them with a user ID, and returns provider authentication links, plus a locally runnable Strata path.^[3][7]

The public repository carries an Apache-2.0 license.^[24] That establishes terms for the repository, not a current hosted-service price or support promise. The reviewed pricing route did not provide a usable integration-platform tariff, so this report does not carry forward an assumed free commercial tier.

Best fit: Teams specifically evaluating Strata's existing implementation or local code, provided they verify the providers, deployment, and support commitment they need. The training-data catalog should not be repurposed as a count of production integration connectors.

Nango

Nango connects managed authentication to deployable action functions. A team enables templates or supplies its own functions, then exposes allowed operations through agent sessions or MCP. Its design is relevant when custom integration behavior matters as much as immediate connector coverage.^[11]

Self-hosting needs a precise distinction. The free edition provides Auth and Proxy with limited observability. Tool calls, syncs, triggers, webhooks, and the MCP server are listed under Cloud or Enterprise self-hosting. Operating the enterprise stack also means managing its services and data stores.^[12] “Self-hostable” does not mean that downloading the free component reproduces the managed product.

Best fit: Engineering teams that want to own integration logic while using managed connection and execution infrastructure. Model persistent connections, compute, and transferred data together; the bill is not simply a tool-call count.^[25]

Paragon

Paragon places ActionKit inside a broader product-integration stack. Customers connect their applications through an embedded experience; an agent or product feature then requests relevant actions.^[26] The current ActionKit documentation also covers triggers, common filter schemas, and reloading customer-specific fields at query time.^[13]

Those details matter for a SaaS product serving customers whose CRM schemas differ. A standard “create contact” action is not sufficient if required custom fields, lookup values, or event subscriptions do not match the customer's actual account. Test those paths alongside ordinary reads and writes.

Best fit: Products needing embedded integrations across agent actions, data synchronization, and workflows. Pricing defines connected users as customer organizations using integrations, so do not estimate it as a simple employee-seat charge.^[27]

Pipedream

Pipedream exposes app-specific MCP tools through Connect and manages user authentication. It supports connecting one's own accounts and building a product that acts for end users. Its docs distinguish free personal MCP use and development mode from embedded production use.^[14]

Workday's February 24, 2026 earnings release confirms that the acquisition closed.^[4] That is a present ownership fact, not a forecast. It neither proves that outside developers will lose support nor establishes that every Workday product feature applies to Connect.

Best fit: Teams whose required applications and actions already fit Pipedream's catalog and connection flow. Check the production allowance for external users separately from execution credits. Repeated reconnects, additional accounts, and workflow compute should be reflected in a representative usage estimate.^[28]

Zapier MCP

Zapier MCP remains an accessible route for connecting supported app actions to an agent. Its product page describes selecting allowed applications and actions and metering an MCP tool call as two Zapier tasks.^[29]

The earlier blanket description of Zapier as unsuitable for embedded end-user authentication is obsolete. White Label documentation describes partner onboarding, a partner-signed identity token, a connection flow, and OAuth credentials for calling the hosted MCP server. The application handles state validation, token storage, refresh, and revocation.^[1] This is a product integration project, not merely copying a personal endpoint.

Best fit: Individuals and teams with suitable existing Zapier actions, and products willing to implement the partner flow. Confirm the commercial terms for that embedded route rather than applying personal-plan assumptions automatically.

StackOne

StackOne offers managed MCP access for employee agents and integration capabilities for embedded products. Its policy surface can restrict tools and inspect or transform data based on configured conditions.^[15][30]

The useful evaluation is whether the exact required provider actions, custom fields, and account boundaries work under the intended policy. For example, a support agent may need read access to customer records while a separate workflow needs a tightly limited write. Verify those behaviors against the provider's final state rather than accepting a connector's response alone.

Best fit: Teams that need centralized tool policy or reusable embedded connections. Gateway seats and embedded-product usage are separate purchase contexts; choose the correct pricing surface before estimating cost.^[16]

Merge Agent Handler

Merge Agent Handler separates identity, selected tools, and data-handling rules through Registered Users, Tool Packs, and its gateway.^[31] Individual and shared credential modes suit different workflows; a shared provider account changes whose permissions and attribution apply.^[17]

Its documentation explicitly says AI Guardrails can fail open when evaluation is uncertain or errors. Use that behavior when designing the surrounding control, rather than assuming every semantic restriction blocks execution.^[18]

Best fit: Governed tool calling where teams want a defined identity and policy model. Resolve the documented commercial discrepancy before budgeting: the pricing page and billing guide currently show different plan names and allowances.^[32][33]

Smithery

Smithery now belongs in this category because Connect provides managed access to MCP servers, beyond indexing them. Its current product and APIs still have a distinct operational model after Arcade's August acquisition.^[2][5]

A concrete behavior to test is partial availability: namespace-wide tool listing can return a successful response containing individual connection errors or omit connections awaiting authentication.^[19] The application should explain missing capabilities instead of letting an agent silently proceed with an incomplete tool set.

Best fit: Applications that benefit from managed MCP connectivity and discovery across servers. Measure protocol traffic, not just completed business actions, because RPC billing includes more than tool execution.^[34]

---

Authentication, Governance, and Deployment

Several independent decisions sit behind the phrase “managed auth.” Keep them separate when reviewing a platform:

Credentials can remain outside the model in a managed connection path while the application still possesses powerful platform tokens. The meaningful question is which components can exercise authority, not merely whether a secret appears in the prompt. Revocation, environment separation, logging, and approval for consequential writes need explicit tests.

Likewise, a source repository, SDK license, customer-managed encryption key, and private deployment solve different problems. The current evidence does not justify reducing every product to a single open-source checkbox or assuming that an unmentioned feature is absent.

Pricing: Compare Billing Units First

Public terms below were reviewed September 16, 2026. Amounts are USD; usage, enterprise contracts, upstream services, and model costs may change the total.

Composio's managed shared OAuth apps have a lower free-call allowance and a different paid rate than bringing an app. Additional routes and features can add charges; do not multiply all calls by the lowest advertised unit price.^[23] Nango rounds each function run to a second, so aggregate runtime alone can underestimate many short executions.^[25]

For a fair trial, record successful tasks, total attempted operations, discovery requests, active users/connections, and any additional compute. Then calculate the cost of the same completed workflow under the applicable plan. A platform with a cheaper call can still need more calls, more implementation effort, or a more expensive support contract.

---

Where Tembo Fits

Tembo runs coding agents in cloud environments with repository context and connected work systems. Its current documentation describes agents using integrations and MCP during both interactive sessions and scheduled runs, including work that spans repositories.^[6][35] Disclosure: Ry Walker is Tembo's CEO and co-founder.

This makes Tembo relevant when the desired result is an engineering outcome: investigate a ticket using code and issue context, prepare a change, and hand it back for review. The integration platform decision becomes relevant when building reusable tool access inside a custom product or extending the agent's available systems.

Tembo's action documentation also covers issue updates, messages, documents, and database queries; its role is broader than simply opening a pull request.^[35] The tradeoff is scope: choosing an engineering execution platform does not provide evidence that it replaces every connector provider's embedded authentication API, catalog, or commercial deployment model. This review does not establish a packaged integration between Tembo and every vendor in the table.

The separate Tembo Agent Studio makes that division concrete for general team agents. It consumes Composio connections and native provider MCP servers, with authorization tied to the acting user and workspace. A manual run uses the caller's connected accounts; schedules and event triggers use their owner's accounts. Choosing Studio therefore still leaves a decision about the underlying connection, provider permissions, and service costs.^[36]

Studio also exposes its own MCP server for inspecting agents and runs, validating definitions, and starting work. That server operates a Studio workspace; it is different from a vendor-neutral connector catalog to embed in any SaaS product. Reads require a viewer role, while run and authoring actions require operator authority and the applicable OAuth scope. It is useful adjacent orchestration context, not an eleventh integration-platform member.^[37]

Selection and Acceptance Tests

These are evaluation starting points derived from the capabilities above, not benchmark rankings. A smaller integration set that reliably completes the required workflow can be preferable to a larger catalog with an unsuitable action or permission model.

Run a trial with two users in the same customer organization and two in different organizations. Connect distinct test data, exercise the intended reads and writes, then try cross-user identifiers and a revoked credential. Verify the actual record at the provider as well as the integration response. Repeat with custom fields, pagination, and an upstream failure.

For consequential writes, define who approves the action and how the application avoids duplicate effects after a timeout. Check that logs explain the acting identity, selected connection, attempted operation, and final outcome. These are proposed acceptance tests; this report did not conduct a comparative hands-on benchmark.

Market Outlook

The verified acquisitions show ownership consolidation, but do not establish a timetable for future deals or universal product convergence.^[4][5] The more useful signals to monitor are concrete: connection portability, policy failure behavior, client compatibility, provider-specific depth, and the commercial cost of operating a real workflow.

Bottom Line

Start with the user and customer boundaries, then verify the exact actions. Compare production setup, failure handling, deployment terms, and cost per completed task before catalog size. For engineering work, evaluate Tembo's execution workflow in the same decision process while keeping its role distinct from the reusable integration infrastructure covered here.

---

Research by Ry Walker • methodology