How many MCP servers are abandoned?

A June 2026 audit of 1,847 MCP servers found 52% abandoned, 31% lightly maintained, and only 17% meeting a reasonable production bar.

Why do MCP tool calls fail silently in agent clients?

When an MCP server updates its protocol version or changes its tool schema, clients that have cached the old schema stop emitting mcpToolCall events with no error. Cursor confirmed this in April 2026 where tool calls worked in 2026.04.14 and silently broke in 2026.04.17.

Why did Perplexity drop MCP?

At the Perplexity Ask 2026 developer conference, Denis Yarats announced moving to direct REST API calls and CLI interfaces, citing MCP authentication friction and tool schema bloat consuming context window space at enterprise scale.

What would an MCP health monitoring layer look like?

It would continuously ping registered MCP servers, version-test their tool schemas against the active MCP spec, flag schema drift to agent clients, and surface a compatibility matrix so developers can see which servers are live before connecting.

What is the business opportunity in MCP reliability tooling?

13,000+ MCP servers are registered but the ecosystem has no SLA or health guarantee. Any agent business that depends on MCP integrations needs a reliability layer. The space is currently wide open with zero dedicated products as of June 2026.

← Back to dashboard

clawsmith.com/signal/mcp-server-ecosystem-52-percent-abandoned-silent-tool-failures

⚠ IssueWide OpenToolLive

MCP server ecosystem is 52% abandoned and tool calls fail silently in agent clients

52% of audited MCP servers are abandoned and 17% are production-ready. Cursor agent CLI silently stopped emitting mcpToolCall events after a version update. Perplexity moved off MCP for internal systems citing authentication friction and tool schema bloat. No registry health check or compatibility layer exists to signal to agents which MCP servers are live.

Product Idea from this Signal

An API that gives MCP server operators real-time health checks, OAuth flow tracing, and client-compatibility scoring across the major AI clients

612 ▲

MCP servers break silently in production: OAuth flows stop working after client updates, 52% of public MCP servers are unreachable, and developers have no surface to know whether their server is healthy for Claude.ai, Cursor, or any other client. There is no runtime observability layer between MCP server authors and the clients consuming them. This API gives MCP server operators continuous health monitoring, per-client OAuth trace logs, and a compatibility score so they can detect and fix failures before users abandon their integration.

mcpai-agentsdeveloper-toolsoauthobservabilityreliability

Competitive79 leadsView Opportunity →