Quick Verdict
UnClick is a hosted MCP layer for tools, connections, and persistent memory. It aims to let Claude, ChatGPT, Cursor, or a custom MCP client discover and call many applications through one setup, with Memory, Connections, AutoPilot orchestration, and XPass checks around the tool catalog. As of July 21, 2026, the official catalog stated 666 apps and 1,580 customer-visible actions. Those are current catalog counts, not evidence that every action was independently tested or that every app has complete coverage.
The product remains in beta and was free at the cutoff. It is useful for low-risk integration experiments, but a single MCP connection should not become an excuse to authorize personal email, drives, repositories, finance, and social accounts all at once. The main risks are connection tokens, tool arguments and response logs, persistent memory, and write tools that can send, modify, delete, publish, or trigger business workflows.
Best For
- MCP application developers comparing SaaS actions and schemas quickly
- Prototype teams using test accounts before committing to production reliability
- Persistent-context agents whose operators can define retention, correction, and deletion
- Internal automation teams with OAuth scope controls, log review, and write approvals
- Not ideal for connecting every personal account, self-hosting requirements, strict SLA needs, or consequential deterministic automation without human review
Key Features
- One MCP entry point for discovering actions across many applications
- Dynamic catalog currently listing 666 apps and 1,580 actions, updated server-side
- Search, tool information, and generic invocation patterns for schema discovery before calls
- Cross-session Memory for facts, identity, and session context, with staleness and sensitive-retention risks
- Connections centralizing third-party authorization and therefore concentrating valuable tokens
- AutoPilot and XPass, described by the vendor as planning, routing, checking, and proof layers, not substitutes for application-side confirmation
Use Cases
- Search the catalog and inspect permissions before connecting an application
- Test read-only operations with isolated, non-sensitive accounts
- Retain user-approved project preferences with view, correction, expiry, and deletion controls
- Draft email, tasks, or record changes and require confirmation before any external write
- Compare UnClick with direct MCP servers, workflow platforms, and official APIs
Pricing
UnClick’s FAQ said it was free during beta at the cutoff, so the canonical pricing is 免费, not freemium. Post-beta pricing, fixed request limits, connector-specific limits, concurrency, Memory capacity, and production SLA were not published. The 666-app and 1,580-action figures describe catalog breadth, not free quotas.
The site links to @unclick/mcp-server on npm, whose package metadata declares MIT. However, the linked GitHub source repository returned 404 during verification. It is accurate to say that the npm package declares MIT, but not that a currently accessible, auditable source repository or reproducible build is available. Nor does the package license automatically cover the hosted backend, connectors, catalog data, or operated service. Teams requiring source review or self-hosting should obtain the corresponding source and build documentation before adoption.
Pros
- One MCP setup can accelerate exploration across a broad app catalog
- Server-side catalog updates reduce per-action package maintenance
- Connections, Memory, and tools support rapid cross-application agent prototypes
- Free beta access lowers the cost of isolated technical evaluation
Cons
- Beta quotas, future price, SLA, and long-term stability are unpublished
- Quality, permissions, and error semantics vary across 666 apps and 1,580 actions
- npm declares MIT while the linked source repository returns 404
- Centralized third-party tokens increase the impact of one compromise
- Persistent Memory and logs can retain stale, incorrect, or sensitive data
- Generic write access can turn model mistakes into consequential external changes
Alternatives
| Tool | Better for | Strength | Limitation |
|---|---|---|---|
| n8n | Teams wanting visual control over connections and approvals | Mature workflows, credentials, and self-hosting | Agent discovery and memory require design work |
| Dify | Agent applications with knowledge and controlled workflows | Centralized app, model, and dataset orchestration | Broad SaaS actions still need plugins or APIs |
| Coze | Fast user-facing bots and workflows | Easy prototyping and channel publishing | External actions and governance need separate review |
| GitHub MCP | Developers needing only GitHub with narrow official scopes | Easier permission and action boundaries | No unified 666-app catalog |
FAQ
What category is UnClick?
It delivers tools, connections, and Memory to clients through Model Context Protocol, so MCP is the correct category.
How many apps and actions does UnClick have?
As of July 21, 2026, the official catalog stated 666 apps and 1,580 customer-visible actions. Catalog content changes, and each action still needs practical verification.
Is UnClick free?
It was free during beta. Future pricing, fixed quotas, and SLA were not published, so permanent free or unlimited usage should not be assumed.
Is UnClick open source?
The npm package declares MIT, but the linked GitHub repository returned 404 during verification. That is insufficient to claim that the full service source is available or the hosted backend is self-hostable.
Should every personal account be connected at once?
No. Begin with test accounts, grant only the minimum OAuth scopes for one task, and maintain token revocation and log-review procedures.
What belongs in persistent Memory?
Only user-approved, useful, non-sensitive facts with correction and deletion controls. Secrets, verification codes, payment, health, identity, employee, and customer-sensitive data should not enter generic long-term memory.
How should write actions be controlled?
Draft first, display the application, account, target, and full change, then require confirmation. Add stronger approval for deletion, publishing, permissions, orders, finance, and employee data, and check application state before retrying a timed-out write.
Bottom Line
UnClick’s proposition is one MCP layer for 666 apps, 1,580 actions, connections, and cross-session Memory. That lowers experimentation cost; it does not remove integration risk. Keep catalog breadth separate from verified quality, the npm MIT declaration separate from the 404 source repository, and technical callability separate from authorization to write. Start with test accounts, read-only actions, no sensitive Memory, and explicit approvals. Expand only after token handling, logs, deletion, quotas, errors, and retry behavior have been validated.