Skip to content
AI & Automation

Raycast MCP vs Zapier MCP: Which Fits a Small Business

Nov 19, 2026

Raycast MCP is a personal Mac tool, and Zapier MCP is the shared-work option

Raycast launched Raycast MCP in its v2.13 macOS release, dated November 18, 2026 according to Raycast (2026), and it lets agents such as Claude, Codex and Cursor use Raycast's own features on a Mac. For a small business the decision turns on whose work the agent does: choose Raycast MCP when one person wants an AI assistant to act on their own Mac, and choose Zapier MCP when a team needs agents to act across shared business apps with access settings and a run history.

TL;DR: Raycast MCP turns the Raycast app on one computer into a local tool server for AI apps, while Zapier MCP is a connection layer that Zapier manages across many business apps. Buy the first for individual desktop work, and use the second, or a reviewed workflow, for anything a team has to audit.

The Model Context Protocol (MCP) is a standard that lets an AI app call tools exposed by another program, and an MCP server is the program that exposes them. Searches for this name also return third-party projects called raycast-mcp-server and RayBridge. Those are community code, not part of this launch, so this article relies on Raycast's own changelog and manual plus the rival and independent sources named below.

Key Takeaways

  • Raycast MCP ships in Raycast's v2.13 macOS release and makes Raycast's Mac features callable by Claude, Codex and Cursor.

  • The Raycast MCP Server is off by default and listens only on your own computer, and the manual says it requires Raycast Pro.

  • Raycast does not prompt for confirmation when a connected app calls a tool, so the review step sits with the connected app.

  • Zapier MCP runs through a connection layer that Zapier manages, with access settings per app and action and a History log of MCP actions.

  • For a team, the approval step is yours to design, because Raycast's pages do not describe team-level approvals for MCP calls.

What Raycast changed in v2.13

The changelog describes Raycast MCP as "a local server that lets agents like Claude, Codex and Cursor use Raycast." You turn it on under Settings → MCP → Raycast MCP Server and then add it to your AI app, as the v2.13 changelog explains. Once it is on, an agent can search Clipboard History, create Snippets and Quicklinks, manage apps and Auto Quit rules, work with browser tabs, history and bookmarks, start Focus sessions, change Raycast settings and remember things across chats. The same changelog says you decide which extensions to share and that everything runs locally on your Mac.

The Raycast manual adds the controls a team will ask about. The server is off by default on each device and listens only on your own computer. Each connected app gets its own token, and each token can be revoked separately. Raycast does not ask for confirmation when a connected app calls a tool. Shared extensions that are disabled, restricted by your organization or unsupported on the platform are unavailable to connected apps, and automations are off by default. The manual also says the Raycast MCP Server is available with Raycast Pro, and that on Mac and Windows other AI apps can use Raycast as an MCP server. The changelog itself is filed under macOS.

Before this release, Raycast sat on the other side of the protocol. The manual describes Raycast connecting to external MCP servers through an Install MCP Server command, using either a local process over standard input and output or a remote HTTP endpoint. The v2.12 changelog already included a fix for MCP servers that stopped while connecting. So the change is one of direction: Raycast used to consume other tools, and now it exposes its own features to other AI apps.

Three gaps matter for a team. The vendor has not published a per-tool approval setting for MCP calls, a run history for those calls, or whether Teams plans include the MCP server.

How we evaluated these tools

This is an informed buyer's guide built from public vendor pages and independent sources. No product was run for it. Every fact comes from a page opened for this article, or it is marked as not published. The weights below are our judgment about what a small B2B team decides on, not vendor data, so a team with a different risk tolerance should change them.

Decision factorWeightWhy it carries this weight
Reach: which tools and data the agent can touch20%Decides whether the tool helps with the work you actually do
Control: approvals and access settings before an action runs20%A wrong action in a shared system is costly to undo
Audit: a run history you can show later15%Needed for review and for knowing what an agent changed
Location: where tool calls run and what leaves the machine15%Shapes your data-handling review
Team sharing and admin controls10%Separates personal use from organization use
Starting cost10%Sets the cost of a pilot
Setup effort10%Sets how fast one person can start
Total100%Weights sum to 100

Reach and control carry the most weight because a tool that touches the wrong system without a review step is the most expensive failure.

Feature matrix on the decisions that matter

Each cell states what the vendor's pages say. "Not published" means the pages reviewed are silent on that point.

Decision factorRaycast MCP (v2.13)Zapier MCPAlfred
Where tool calls runLocal, on your MacConnection layer that Zapier manages; the pages do not say where calls executeNo MCP on its site
AI clients namedClaude, Codex, CursorClaude, Claude Code, ChatGPT, Cursor, Codex, and any client that supports Streamable HTTPNot mentioned
What the agent can reachClipboard, snippets, quicklinks, apps, browser tabs and history, Focus, settings, memory, shared extensionsApps and actions you enable from a large catalogHotkeys, keywords and workflows
Approval before a tool runsNo per-call prompt; the connected app decidesOnboarding asks you to approve the first tool run; access is set per app and actionNot applicable
Run historyNot publishedEvery MCP action appears in HistoryNot applicable
Team admin controlsEnterprise AI Control Center, not MCP-specificEnterprise app and action access; IT can enable or disable MCP per workspaceNot described
Starting planPro, which the manual names for the MCP serverFree tier, then paid tiersPaid single license

Read across the rows. Raycast MCP differs from Zapier MCP on where calls run and on whether a per-call prompt exists. Zapier differs on run history and administrator controls. Alfred is included because it is the other Mac launcher a buyer is likely to weigh. Its pages do not mention AI or MCP, so it cannot be scored on agent features from its own material.

Pricing and cost to start

Pricing checked November 18, 2026.

Raycast Pro costs $10 per month or $96 per year, and Teams Pro costs $15 per user per month or $144 per user per year, according to Raycast (2026). The pricing page does not mention the MCP server, while the manual names Pro as the plan that includes it, so confirm the plan with Raycast before a team purchase.

On Zapier, the Free plan includes 100 tasks per month, the first paid Professional tier with 750 tasks costs $29.99 per month, and the Team entry tier with 2,000 tasks costs $103.50 per month, according to Zapier (2026). Zapier also says MCP is included in every Zapier plan, on its MCP page.

On Alfred, the single license costs £34, according to Alfred (2026). The page does not say whether that is a one-time purchase or a subscription.

VendorPlanPrice shownBillingWhat the page says it includes
RaycastPro$10 per monthMonthly; $96 per yearRaycast AI; manual says the MCP server needs Pro
RaycastTeams Pro$15 per user per monthMonthly; $144 per user per yearAI features; MCP inclusion not stated
RaycastEnterpriseQuote-basedQuote-basedAI Control Center and admin controls
ZapierFreeFreeFree100 tasks per month, basic MCP
ZapierProfessional (first paid tier)$29.99 per monthMonthly price shown750 tasks per month
ZapierTeam (entry tier)$103.50 per monthMonthly price shown2,000 tasks per month
ZapierEnterpriseQuote-basedQuote-basedCustom annual task limit
AlfredSingle License£34Not stated on pageAlfred 6 upgrade listed

Each price in the table is cited in the paragraphs above.

Here is an illustrative case, not a measured result. Assume a five-person team where agents make 40 successful Zapier tool calls a week. At 2 tasks per successful call, that is 80 tasks a week, or 320 tasks in a four-week month, which fits inside the 2,000-task Team entry tier, the plan a five-person team sharing one account would need because Professional is a single-user plan. The setup step matters as much as the price. The Zapier docs name discover_zapier_actions to find an app and enable_zapier_action to add it, so an agent can enable an action during the same request, and the team should decide in advance who may enable actions.

Week of agent useSuccessful callsTasks per callTasks used
Week 140280
Week 240280
Week 340280
Week 440280
Month total1602320

The same habit of watching unit costs applies elsewhere, as the vendor price comparison guide for restaurant purchasing shows.

Vendor profiles: best fit, limits and evidence

Raycast MCP

Best fit: one person, or a small group of Mac users, who want Claude, Codex or Cursor to work with their own clipboard, snippets, quicklinks, app controls, browser tabs and Focus sessions. The launch is built around personal productivity, not a shared operations queue.

Limitations: the server listens only on one computer, so there is no shared administration for it. The manual places it under Raycast Pro. Raycast does not prompt before a connected app calls a tool. The vendor has not published run history for MCP calls or whether Teams plans include the server.

Implementation: turn it on under Settings → MCP → Raycast MCP Server, read the example prompts in the same place, choose which extensions to share, and add the server to each AI app. Each app gets its own revocable token, so issue tokens per person rather than sharing one. Review each shared extension before enabling it, because connected apps can call tools without a prompt.

Primary evidence: the v2.13 changelog and the Raycast manual.

Zapier MCP

Best fit: teams that need agents to act across many business apps and want shared access rules, IT switches and a history of what ran. Zapier's clients page lists Claude, Claude Code, ChatGPT, Cursor and Codex, and says the setup works with any client that supports Streamable HTTP.

Each successful Zapier tool call uses 2 tasks according to Zapier (2026), and failed calls do not count. Task volume is therefore the main cost driver to watch.

Limitations: usage is metered, so heavy agent use can outgrow a small tier. Zapier's pages give two different action counts, so ask for a live list of the actions your apps expose. The onboarding asks you to approve the first tool run, and the pages reviewed do not describe approval for each later call. They also do not say where tool calls execute.

Implementation: add Zapier as a connector in your AI client and sign in when prompted. The docs say Zapier creates your default server on first connection, and an admin may need to approve the connector if your organization manages the client. The docs name enable_zapier_action as the call that adds an action to the server, and discover_zapier_actions to find an app's actions. Decide who may enable actions before the first run.

Primary evidence: the Zapier MCP docs and the Zapier MCP page.

Alfred

Best fit: Mac users who want hotkeys, keywords and workflows that need no coding, and who want to import workflows shared by the community. Its page highlights workflows, Mac control, 1Password bookmark access and settings sync across Macs.

Limitations: the page does not mention AI, MCP or Model Context Protocol, so Alfred cannot be compared on agent features from the vendor's own material. The page says some features require the Powerpack.

Implementation: buy the Powerpack license that fits your needs. The page says buying now includes a free upgrade to Alfred 6, and it lists a higher tier with free lifetime upgrades.

Primary evidence: the Alfred Powerpack page.

Security and approvals: where the risk sits

Local tools carry their own risk. A Backslash Security analysis reported by IT Pro estimated that around half of the 15,000-plus MCP servers in existence were dangerously misconfigured or carelessly built, according to IT Pro (2025). That figure predates this launch and covers MCP servers in general, not Raycast or Zapier.

Security concerns: top MCP adoption obstacle at 64% according to Stacklok (2026), among software-sector respondents to its survey of senior technical leaders.

Our read: Raycast's local-only design, as the manual describes it, reduces the network exposure that the IT Pro analysis describes, because the server does not listen beyond your computer. It does not remove the need to vet each shared extension and connected app. Zapier's access settings and run history address a different risk, the team's record of what ran, and they come with dependence on a service that Zapier manages. For another enterprise example of an MCP-connected launch, see the iManage and Thomson Reuters explainer.

Where stitching it together fits

For many small teams the real alternative is stitching the same work together in Zapier, Make or n8n, or building it in-house. Those tools can keep run histories, retries, error branches and audit evidence when they are configured. The buyer, though, has to design and own observability, idempotency (so a retry does not repeat an action), escalation paths, access controls and ongoing maintenance. This article did not review the Make or n8n pages, so it does not price or rank them.

US Tech Automations can configure a stitched design of this kind. The two descriptions below are proposed, configurable workflows, not live deployments.

The first proposed workflow starts when a deal in the CRM moves to a "Proposal sent" stage. An agent reads the call notes from a CRM export or API and drafts a follow-up email plus a task list. The output is a draft and a task waiting in a review queue for the account owner. The prerequisites are API access or a scheduled export from the CRM and a named owner for the queue. The human review point is the account owner, who approves before any email leaves.

The second proposed workflow starts each week, when a scheduled export of open support tickets arrives. The agent groups tickets by product area and flags those that breach an age threshold the operations lead sets. The output is a summary posted to one Slack channel. The prerequisites are ticket export access, a Slack app with permission to post, and a written threshold. The review point is the operations lead, who decides what escalates. No ticket is changed automatically. For a related Slack hand-off setup, see the monday.com to Slack automation guide.

Compared with Zapier MCP's History log, a proposed US Tech Automations design would put the run log, retry rules and escalation path into the design from the start. It would still depend on your export access and on a reviewer who owns the output. Raycast MCP has no equivalent design step because it is a tool one person drives, not a shared process.

When NOT to use US Tech Automations: a simpler tool wins when one person wants shortcuts on their own Mac, where Raycast MCP alone is enough. It also wins when one app-to-app connection with no approval step will do, where a native automation or Zapier MCP is quicker to set up. And if you lack API or export access or a named reviewer, do not build a workflow yet, because it would only create output no one owns.

Who this is for

Who this is for: a founder, operations lead or engineering lead at a small B2B software company who wants to know this month whether an AI agent can safely act on their tools, and who can name one owner for approvals. Raycast MCP suits the individual on a Mac. Zapier MCP suits a team with shared apps. Neither suits a team with no reviewer.

Red flags: you need approvals before any tool call runs; you need a team-wide history you can show a customer or auditor; your work sits mainly in cloud systems rather than on the desktop.

Decision checklist

  1. List the apps and data the agent must touch this quarter, and mark which are shared.

  2. Name one person who approves any action that changes a shared system.

  3. Ask Raycast in writing whether your plan includes the MCP server, and whether Teams plans do.

  4. For Zapier, multiply your expected calls per week by 2 tasks per successful call, then compare the monthly total with the task limit of the tier that fits your team size, since Professional is a single-user plan and a shared team account needs the Team tier.

  5. Decide in advance who may enable Zapier actions, since an agent can enable one itself.

  6. Keep a named owner for run history, even when the vendor provides one.

Signal vs Speculation

Signal (stated by vendors or sources):

  • Raycast MCP is part of the v2.13 release described in Raycast's changelog, and the manual says it requires Raycast Pro.

  • The manual says the server is off by default, listens only on your computer, and gives each connected app its own token.

  • Raycast's pages have not published run history for MCP calls or MCP eligibility for Teams plans.

  • Zapier says MCP is included in every plan and that every MCP action appears in History.

  • Alfred's site does not mention AI or MCP.

  • Stacklok's 2026 software-sector survey reports security concerns as the top MCP adoption obstacle at 64%.

Speculation (our read, not vendor fact):

  • For one operator who wants shortcuts on a Mac, Raycast MCP is the simpler start, because the approval step lives in the connected app and the tool runs on the machine.

  • For a team with shared apps, Zapier MCP's access settings and run history make it the safer default, though its pages do not show how well those controls work in practice.

  • The more an agent can change, the more the review step matters, and with either product the buyer owns that step.

Frequently asked questions

Does Raycast MCP work with Claude, Codex and Cursor?

Yes, the v2.13 changelog names those three as examples of agents that can use Raycast once the server is on. Other MCP clients may also connect, but the changelog does not list them.

Does Raycast MCP need a paid plan?

Yes, the Raycast manual says the Raycast MCP Server is available with Raycast Pro, which is the plan priced above. The pricing page does not mention MCP, so confirm the plan in writing before a team buys.

Does Raycast MCP send data to the cloud?

The vendor says everything runs locally on your Mac. Whatever the connected AI app sends to its own model service is governed by that app's terms, not by Raycast.

Is Zapier MCP the same product as Raycast MCP?

No. Zapier MCP is a connection layer that Zapier manages across business apps, while Raycast MCP exposes features of the Raycast app on your own computer.

Should a small business act on this launch this month?

Act this month only if you can name an owner, confirm your Raycast plan, and decide which extensions to share. Otherwise wait, because the vendor has not published team-level controls for MCP calls.

Does Alfred offer MCP?

Its website does not mention MCP or Model Context Protocol, so do not assume support. Ask Alfred directly before you plan around it.

Bottom line for this month

Raycast MCP is a narrow launch with a clear job: it lets agents on your Mac use your Raycast features, and it puts the approval step in the connected app. If your question is about shared business apps, compare it with Zapier MCP on approvals, history and task costs before you choose. If you need a reviewed, team-owned workflow, see how US Tech Automations configures this for a proposed design with its prerequisites and review points named up front.

About the Author

Garrett Mullins
Garrett Mullins
Workflow Specialist

Helping businesses leverage automation for operational efficiency.