Unified MCP vs. Merge Agent Handler: which MCP server should your agent use? (2026)

Choose Merge Agent Handler if your agent acts mostly in productivity and SaaS tools and you need a gateway that can redact or block what the agent sends. Choose Unified MCP if your agent needs HR, payroll, recruiting, accounting or CRM data in one normalized shape across vendors, with no tool results kept in logs. Both are hosted MCP servers that call third-party systems live, and at list prices as of 8 October 2026 a tool call costs about $0.04 on Agent Handler Pro and about $0.001 on Unified.to Grow (our calculation, assuming each entry plan's included allowance is fully used; Unified.to's allowance also covers your other API usage).

Agent Handler is a separate product from Merge Unified, Merge's unified API. Merge's own guidance splits them: "For data sync, integration, or analytics: Unified API. For agents that take actions: Agent Handler. Some teams use both." (Merge FAQ) Using both means two vendors, two credential stores and two bills, so price both before committing. This post compares Agent Handler with Unified MCP. For Merge Unified against Unified.to's API, see Merge vs. Unified.to. For Unified MCP against another agent tool-calling product, see Unified MCP vs Paragon MCP Server.

How this comparison was built. This post was written by Unified.to. Merge facts come from Merge's public Agent Handler docs, pricing page and product page, retrieved 8 October 2026. We didn't test Agent Handler hands-on. Where Merge's pages disagree with each other, we say so.

How do Unified MCP and Agent Handler compare at a glance?

Unified MCP returns normalized objects across 1997 integrations and bills a tool call as one API request; Merge Agent Handler returns each connector's own tool schema across 216 Merge Connectors and bills a tool call as one credit. Values below are as of 8 October 2026.

• Both call the source live. Merge says Agent Handler tool calls "aren't cached server-side", and Unified MCP routes every call to the source API at request time.
• The data model is the real split. Agent Handler tools follow each vendor's schema; Unified MCP tools return one normalized shape per category, with raw fields and vendor endpoints available when the model doesn't cover something.
• The catalogues lean different ways. Agent Handler's is built around productivity and SaaS tools such as Google Workspace, Microsoft 365, Slack and Teams. Unified MCP includes HR, payroll and recruiting systems that Agent Handler's connector index doesn't list as of 8 October 2026.
• Agent Handler has a free plan; Unified.to has a 30-day trial on Grow and Pro. Agent Handler's Free plan includes 2,000 tool calls a month and stops at the limit. Past the trial, Unified.to's price also covers change events that carry the changed record on supported integrations; Agent Handler emits three webhook event types.

What is Merge Agent Handler?

Merge Agent Handler is a hosted MCP server that routes an agent's tool calls to third-party tools with each end user's credentials, scanned by security rules and logged. Merge describes it as "the tool-calling platform for AI agents" and says it is generally available (overview).

The pieces, in Merge's terms:

• Connectors. A Merge Connector is an integration Merge builds and maintains, with its own tools such as slack__post_message or jira__create_issue. Generic connectors are MCP servers that vendors build and host, which Agent Handler passes calls through to. You can also register your own remote MCP server.
• Tool Packs. A named bundle of connectors and the specific tools one agent can see. "Most Connectors expose 30–80 tools; you usually want a subset." A connectors query parameter narrows one connection further, and the search_tools meta-tool and Merge CLI let an agent find tools by intent.
• Registered Users. The unit of isolation: every credential, tool call and audit record belongs to one. End users connect their accounts through Merge's Link component or a Magic Link.
• Security Gateway. Rules that allow, redact or block what an agent sends before the call reaches the third party, plus AI guardrails written in plain language.
• Agent Handler for Employees. A second setup in which IT gives its own employees one MCP URL, signed in through the company's identity provider, with skills and per-role tool access.

Agent Handler shares Merge Unified's Link flow and end-user credential store, but not its Common Models or sync layer. None of Agent Handler's pages mention Common Models.

Merge also publishes merge-mcp, a separate MCP package for Merge Unified that runs locally with a Merge API key and one Linked Account token. It isn't part of Agent Handler.

What is Unified MCP?

Unified MCP is a hosted MCP server on top of Unified.to's unified API, which routes every call to the source system at request time and stores no customer records.

• Hosts per region: US mcp-api.unified.to/mcp, EU mcp-api-eu.unified.to/mcp and AU api-au.unified.to/mcp. Each workspace runs in one region: the US on every plan, or the EU or AU on Pro and Scale. A credential works only on its own region's host, and each region sees only its own workspaces. On Scale, Multi-Region Sync keeps your account configuration consistent across regions.
• Protocol: Streamable HTTP. Unified MCP supports MCP revisions 2024-11-05 through 2025-11-25, plus the stateless MCP 2026-07-28 revision in production.
• Tools: normalized objects (Contact, Deal, Candidate, Employee, Invoice, Payslip, Payroll and others) that return the same shape for every integration in a category and write back through the same objects where the integration supports it, plus each vendor's own endpoints as tools with include_external_tools. 22,566 tools in all. That total is the catalogue across all integrations; each connection sees only its own tools: one HubSpot connection carried 69 tools, 30 of them CRM, on 12 October 2026.
• Five ways to authenticate (as of 26 October 2026): a workspace API key with a connection ID; a workspace API key alone, for Core mode; enterprise-managed authentication, enabled per workspace on request, where the customer's identity provider issues a token whose scopes decide which tools the agent can see; a connection-scoped token that grants one connection and keeps your workspace API key on your server; or OAuth sign-in, where an MCP client such as a Claude custom connector signs in to Unified.to itself and gets Core mode tools. Credentials go in the Authorization or x-api-key header.
• Tool controls: tools and permissions limit the tool list, defer_tools defers loading, aliases adds matching hints, and hide_sensitive removes personal data such as names, emails, phone numbers, addresses, date of birth, SSN/SIN and compensation, plus credential fields, from results before they reach the model. Tool lists are available formatted for OpenAI, Gemini, Cohere, Grok and Groq; for Anthropic's Messages API, use the OpenAI format and rename parameters to input_schema, or use Anthropic's MCP connector. See the MCP API endpoints docs.
• Core mode for coding agents: a workspace API key without a connection, or OAuth sign-in from an MCP client, gives management tools for connections, webhooks, API-call logs, notifications, secrets managers, Agent Skills and docs search. It doesn't read or write end-customer data. It installs in Claude Code, Codex, Cursor and VS Code. Unified.to's Agent Skills give coding agents a SKILL.md for each Unified.to API, plus skills for registering your own OAuth app with each vendor, and the Claude Code plugin adds the Core MCP server in the same install.
• Billing and limits: each tool call, and each MCP Events poll, is one API request on your plan; tool calls count against the workspace rate limit, and each delivered MCP event counts like a webhook delivery.

For why an agent might use one normalized MCP server rather than each vendor's own, see why use a unified MCP instead of connecting directly to vendor MCPs.

What does an agent get back from each?

On Unified MCP, a tool returns the same object shape for every integration in its category; on Merge Agent Handler, each connector returns its own vendor's shape. That is the most consequential difference for an agent that works across several systems. The normalized objects are typed: an Invoice, for example, has 40 top-level fields, a 10-value status, and line items coded with project_id and is_billable.

Ask both for a list of contacts from HubSpot and from Salesforce. On Unified MCP, crm_contact_list returns one Contact shape for both, and for any of the 270 CRM integrations. On Agent Handler, HubSpot and Salesforce are separate connectors with separate tools (78 and 77 listed tools respectively as of 8 October 2026), each with its own names and schema, so reconciling them falls to the agent's prompt or your code.

An agent that needs normalized HR, payroll or accounting data through Merge needs two products, billed two ways: Agent Handler per tool call and Merge Unified per Linked Account. On Unified.to, normalized data and agent tool calls come through one MCP server on one plan.

Per-vendor tools have advantages. They follow each vendor's API closely, and Agent Handler's 8,448 listed tools include actions no normalized model would cover. Unified MCP reaches vendor-specific actions through include_external_tools, which adds each vendor's own endpoints as tools alongside the normalized ones. The trade-off is the same one covered in should AI agents use a unified API or native APIs.

Which systems can each reach?

Merge Agent Handler's catalogue is built around productivity and SaaS tools, with 8,448 listed tools across 216 Merge Connectors as of 8 October 2026. Unified MCP covers 1997 integrations across 38 categories, including ten HR and recruiting systems that Agent Handler's index doesn't list.

Agent Handler. As of 8 October 2026, its connector index lists 216 Merge Connectors and 113 generic MCP servers (Merge Connectors, generic servers). The Merge Connectors include Gmail, Google Drive, Docs, Sheets and Calendar, Outlook, OneDrive, SharePoint, Teams, Slack, Discord, WhatsApp, Salesforce, HubSpot, Jira, Zendesk, Stripe, NetSuite, QuickBooks Online and Xero, and in HR and recruiting, BambooHR, HiBob, Workday, UKG Pro, Greenhouse, Ashby and SmartRecruiters.

Three caveats come with those numbers:

• Merge's own pages count differently: "190+ MCP-ready third-party connectors" on the overview, "hundreds of connectors" on the connectors page and "thousands of enterprise tools" on the product page.
• The 113 generic servers are vendor-built and list no tools. Merge says the vendor "can change that surface without warning."
• Docs tool lists trail Merge's marketing pages. Slack lists 50 tools in the docs and 78 on merge.dev.

Unified MCP. 1997 integrations across 38 categories, including HR & Directory (808), ATS (179), CRM (270) and Accounting (319). Ten systems an HR or recruiting agent needs aren't in Agent Handler's connector index as of 8 October 2026: ADP, Rippling, Gusto, Paychex, Paycor, Paylocity, Dayforce, Personio, Lever and iCIMS. All are on Unified.to, and pay runs and pay codes are normalized objects on payroll systems including Paychex and Paycor. It also has normalized Messaging (115), Calendar (85) and Task (145) categories, covering Gmail and Outlook messages, Google Calendar and Microsoft Teams events, and Asana tasks, among others. The full list is at unified.to/integrations.

If your agent needs many vendor-specific actions in email, documents, chat and project tools (Agent Handler's docs list 50 Google Sheets tools as of 8 October 2026), Agent Handler's Merge-built connectors are the closer fit. If it needs employee, payroll, candidate, accounting or CRM data in one shape, Unified MCP is.

What does each keep about your customers' data?

Neither Merge Agent Handler nor Unified MCP keeps a synced copy of source data, but Agent Handler logs each tool call's arguments and result, and Unified.to logs metadata only.

Agent Handler. Each tool call log row includes "Arguments - exactly what the agent passed in, post-redaction" and "Result - exactly what was returned, post-redaction" (tool call logs). As of 8 October 2026, logs, the audit trail, credentials and configuration are stored on AWS US-East, encrypted at rest with AES-256, with logs and the audit trail kept for "90 days by default." Files a tool returns are staged by Agent Handler and served by a download URL valid for an hour.

Unified MCP. No customer records are stored. Unified.to's API-call logs keep status, timestamps, path, method, size and an error description, with no request or response bodies; its privacy policy states that logs "contain only operational metadata" and that it does "not store, persist, or maintain end-customer data transmitted through our APIs." As of 8 October 2026, retention is 60 days on Grow and Pro and 365 days on Scale. What persists beyond logs is connection credentials and minimal operational metadata, both encrypted at rest (the metadata with AES-256), and on Pro and Scale credentials can live in your own secrets manager instead.

Getting logs into your own monitoring differs too. Agent Handler streams tool call and API request logs to a SIEM on Enterprise only, and its log endpoints serve the last 30 days (SIEM streaming). Unified.to ships logs to your own Datadog, Grafana or ClickHouse on Pro and Scale.

How does an agent find out that something changed?

Unified.to sends change events from supported source systems through native and virtual webhooks; Merge Agent Handler forwards webhooks that some connectors send, unnormalized, alongside two account events.

As of 8 October 2026, Agent Handler emits three webhook event types, and Merge calls the table "the complete set": REGISTERED_USER_CREATED, REGISTERED_USER_DELETED and INBOUND_WEBHOOK_RECEIVED (webhooks). INBOUND_WEBHOOK_RECEIVED forwards a subscribed connector's own webhook (Merge names Slack, Jira and GitHub) in the vendor's format, passthrough only. Where a connector sends none, an agent that needs to know a candidate moved stage or an invoice was paid has to ask, by calling a tool.

Unified.to delivers change events on supported integrations: native webhooks where the source sends them, and virtual webhooks where it doesn't, checked on an interval you set, down to one minute on paid plans. Where supported, events carry the changed records, so there's usually no follow-up call, and checks that find nothing aren't billed. include_all loads existing records before changes start, and virtual webhooks detect deletions on ten integrations, including BambooHR, Box, Google Drive and OneDrive. An agent can act on the event rather than polling with tool calls, each of which would spend a credit on Agent Handler. As of 26 October 2026, a connection's Unified MCP server also supports MCP Events, a draft extension of the MCP specification: the agent's client can poll for changes with a cursor or subscribe an HTTPS webhook, signed per Standard Webhooks, for event types such as crm_contact.created. Events aren't offered in Core mode.

How do the security controls compare?

Merge Agent Handler offers more ways to inspect and block what an agent sends; Unified MCP removes personal data from what comes back and limits which tools an agent sees.

Agent Handler's Security Gateway (docs):

• Entity rules detect personal data such as names, emails, phone numbers, SSNs and card numbers on a Presidio-based stack. Each rule can allow, redact or block, and redaction replaces the value with a placeholder such as [REDACTED:EMAIL] before the third party sees it.
• Custom regex rules work the same way.
• AI guardrails are rules written as a sentence. They can block or log, not redact, and run both before the call and after the third party replies. When evaluation itself fails, "the call continues," which Merge describes as "the opposite way from a fail-closed design" (guardrails).
• As of 8 October 2026, Merge's pages disagree on whether entity and regex rules also scan tool results. The Security Gateway page says it "scans tool arguments" and that guardrails are the rule type that screens responses; the entity and regex pages say they scan "tool inputs and outputs."

Unified MCP's controls:

• hide_sensitive removes personal data, including names, emails, phone numbers, addresses, date of birth, SSN/SIN and compensation, plus credential fields, from results before they reach the model. It's off by default and set per URL or call.
• tools and permissions decide which tools an agent sees at all.
• With enterprise-managed authentication, enabled per workspace on request, the customer's identity provider issues scoped tokens, and a tool the token couldn't use is never offered.
• A connection-scoped token grants one connection and never carries your workspace API key; it doesn't expire, so you revoke it by deleting the connection. A connection-scoped session only shows the agent tools for that connection's integration.
• A workspace IP allowlist applies to workspace API keys and enterprise-managed tokens over MCP: a call from an address not on the list gets a 403. Connection-scoped and OAuth tokens aren't held to it.

If you need to redact what an agent sends to a third party, Agent Handler has the more developed tooling. If you need personal data kept out of the model's context, Unified MCP's hide_sensitive removes it from results on request.

How do authorization and credentials work?

Merge Agent Handler uses Merge's OAuth apps by default, so end users authorize Merge; Unified.to has you use your own OAuth app in production, so they authorize you.

Agent Handler. "By default, Agent Handler uses Merge's OAuth apps for each Connector." You can bring your own, and a few connectors require it. A callback on your own domain "works only alongside your own OAuth app" (application credentials). Calls carry an access key in the Authorization header, and the URL names the Tool Pack and the Registered User. Merge recommends minting user-scoped keys that can't call tools as anyone else.

Unified MCP. In production you register your own OAuth app with each vendor, so the consent screen your customer sees carries your name. Unified.to publishes OAuth app-registration skills that walk a coding agent through each vendor's developer portal. The authorization component ships for React, Angular, Vue, Svelte and plain JavaScript, and on Pro and Scale the authorization page can sit on your own domain (US and EU regions). Because the OAuth apps are yours, the API returns each connection's tokens, and tokens issued to your app keep refreshing if you leave.

Merge's default gets you live faster on connectors where you haven't registered an app. Unified.to's model puts your brand on the consent screen and keeps the grant with you.

If you're moving a unified API integration from Merge or another vendor, Unified.to's migration guides map authorization, API calls, pagination, webhooks and passthrough, and cover re-authorizing customers. Only tokens issued to an OAuth app you own keep refreshing after you switch.

Where does each run?

Unified MCP and Merge Agent Handler are both hosted services, with on-prem available only on Enterprise. Unified MCP has US, EU and AU hosts, and each workspace runs in one of them: the US on every plan, or the EU or AU on Pro and Scale; as of 8 October 2026, Agent Handler runs in US-East on standard plans, with EU and APAC residency on Enterprise contracts.

Merge's FAQ: "Standard plans run on Agent Handler's hosted infrastructure (US-East). Enterprise contracts can include single-tenant instances, regional residency (EU, APAC), and VPC peering." Its pricing page lists on-prem deployment on Enterprise. Unified.to's Enterprise plan includes single tenant, private cloud, dedicated cloud and on-prem.

On the protocol, Agent Handler negotiates MCP revisions from 2024-11-05 to 2025-11-25, and connects to custom MCP servers on 2024-11-05 and 2025-03-26 only. Unified MCP supports revisions 2024-11-05 through 2025-11-25, plus the stateless 2026-07-28 revision, which drops the protocol-level session so MCP requests route through standard gateways and load balancers like any other HTTP request.

How does pricing compare?

Merge Agent Handler and Unified MCP both bill per tool call, and on list prices as of 8 October 2026, assuming each entry plan's full included allowance is used, a Unified.to request costs about $0.001 against about $0.04 for an Agent Handler credit, by our calculation. Published prices as of 8 October 2026:

On Agent Handler, "Every tools/call against the MCP endpoint counts as one, whether it succeeds or fails. Failed calls and blocked-by-rule calls count," and test-key usage counts the same as production. Listing tools doesn't count.

The arithmetic on those list prices: $1,000 buys 25,000 tool calls on Agent Handler Pro, about $0.04 each. $750 buys 750,000 API requests on Unified.to Grow, about $0.001 each. Two qualifications. Unified.to's request allowance covers all your API usage, not only MCP tool calls, so model your total. And the number of calls an agent needs for a task can differ between platforms.

For prototypes under 2,000 calls a month, Agent Handler's Free plan costs nothing. Past that, the per-call difference dominates. What the difference buys on a paid Unified.to plan: one normalized shape per category, and tool results kept out of logs, where Agent Handler logs each call's arguments and result, after redaction, for 90 days by default.

Which should you choose?

Choose by the systems your agent acts in and what you need kept out of logs. Values as of 8 October 2026.

When should you choose Agent Handler?

Choose Merge Agent Handler when your agent acts mostly in productivity and SaaS tools (Gmail, Outlook, Google Drive, SharePoint, Slack, Teams, Jira) and needs many vendor-specific actions in them. Beyond the rows in the table above, it also fits when:

• You want to plug vendor-hosted or your own remote MCP servers in behind one gateway, with the same credentials and audit trail
• You want to go live on connectors before registering your own OAuth apps
• You're giving your own employees AI tools, with SSO, SCIM and per-role access
• You want agent ergonomics built in: tool search, a CLI for coding agents, per-pack tool description overrides and file transfers up to 100 MB
• Named production adoption matters: Merge's product page quotes Perplexity on using Agent Handler

When should you choose Unified MCP?

Choose Unified MCP when your agent needs HR, payroll, recruiting, accounting or CRM data in one shape across vendors, including systems such as ADP, Rippling, Paychex, Lever and iCIMS. Beyond the rows in the table above, it also fits when:

• You want personal data removed from results before they reach the model
• Your agent should react to changes in one normalized event format rather than polling with tool calls
• You want your brand on the OAuth consent screen and the tokens to stay yours if you leave
• Your clients or gateways need the stateless 2026-07-28 MCP revision
• Your coding agents set up and manage integrations themselves through Core mode and the OAuth app-registration skills

Frequently asked questions

Which AI clients work with each?

Both work with MCP clients that support Streamable HTTP and can send the server's credentials. Agent Handler requires an access key in the Authorization header ("Tokens are not URL parameters"), documents setup for Microsoft Copilot Studio, and offers the Merge CLI for coding agents. Unified MCP takes credentials in the Authorization or x-api-key header, lets MCP clients that support OAuth, such as a Claude custom connector, sign in to Core mode without an API key, formats tool lists for OpenAI, Gemini, Cohere, Grok and Groq (Anthropic uses the OpenAI format with input_schema), and its Core mode installs in Claude Code, Codex, Cursor and VS Code.

For how Unified MCP and Merge Agent Handler compare with eight other MCP servers, see Best MCP Servers for Connecting AI Agents to SaaS APIs.

Start a free trial or book a demo to see Unified MCP work with your agent.

Written by Mallory Greene.

About the author: Mallory Greene is Head of Marketing at Unified.to_. She writes about integration infrastructure, unified APIs, and MCP for technical teams. Based in Toronto._