The Model Context Protocol (MCP) is fast becoming the go-to standard for AI applications needing to connect with external tools. As more agents start making calls through MCP, SaaS teams are facing challenges that go beyond simple 'is the service down?' questions. Instead, they're grappling with more nuanced issues like 'did the agent pick the wrong tool?' or 'did the model pass incorrect parameters?' This is precisely the gap Spanly aims to fill, positioning itself as an observability and monitoring platform specifically for MCP servers.
From what we can gather on their website, Spanly emphasizes a 'drop-in' integration experience. You won't need to tweak your core business logic; instead, you can hook into your existing MCP servers using either a CLI or an SDK. It monitors a comprehensive set of metrics crucial for production-grade services, including error rates, session traces, latency, client analysis, and deployment alerts. This covers pretty much all the key indicators you'd want to keep an eye on.
Beyond Basic Health Checks
Traditional monitoring tools often focus on metrics like p95 latency or HTTP error codes. However, problems within MCP servers tend to run much deeper. Spanly's interface, as shown on their site, categorizes issues into types like 'tool poisoning,' 'schema correctness,' and 'runtime security.' Think about scenarios such as prompt injection appearing in tool outputs, sensitive keys being returned in tool parameters, tool names confusing agents into selecting the wrong command, or schemas accepting malformed arguments. These are not the kinds of insights a standard Application Performance Monitoring (APM) tool can readily provide.
The company highlights that its scanner has already analyzed a significant number of production MCP servers (their site noted '11,127 MCP servers scanned' on the day of review). This extensive, large-scale scanning data allows it to 'know what problems look like,' which is a key differentiator from more generic monitoring solutions.
From Detection to Automated Remediation
What's even more practical is that Spanly doesn't just report issues; it actively tries to suggest patches. Screenshots on their website illustrate how it can propose modifications like renaming a tool, tightening a schema, or trimming outputs. Crucially, it supports an A/B testing approach, first validating these fixes in 10-20% of live sessions before gradually rolling them out to all traffic once effectiveness is confirmed. This 'small-scale validation before full deployment' strategy is a pragmatic move for any production environment.
Furthermore, it categorizes these remediation suggestions by source – some might require an upstream update, while others can be applied directly. This means teams can avoid deep dives into MCP protocol specifics, saving considerable debugging time.
Integration and Data Residency
Spanly isn't looking to replace your existing monitoring stack like Datadog, Sentry, or New Relic. Instead, it's designed to complement them, acting as a specialized probe for the MCP layer. If you're already using these tools, Spanly can slot right in. Additionally, it offers data residency options in both the US and the EU, which is a significant plus for enterprises with strict data compliance requirements.
For deployment, Spanly provides both CLI scripts and an SDK, boasting a 30-second free scan. While a free tier is confirmed, the full pricing structure isn't entirely public, so you'll need to check their official website for complete details.
Who Benefits Most?
If your team is running MCP servers in production, serving various AI agent clients, Spanly could be a valuable addition to your monitoring strategy. It particularly shines in those frustrating situations where 'the service isn't down, but the agent just isn't behaving correctly' – problems that often lie not in HTTP status codes, but in how tools are exposed.
Another common scenario involves an MCP server needing to serve multiple clients like Claude and Cursor simultaneously. Different clients might interact with tools in subtly different ways. The phrase on their website, 'Works in Claude, Broken in Cursor,' perfectly captures this pain point. Using a unified scanning and monitoring solution to surface these issues proactively is far smarter than waiting for customer complaints.
Of course, Spanly is still a relatively new player, and detailed technical insights into its detection mechanisms are somewhat limited. The accuracy of its scanner will also benefit from more real-world deployments. It's a good idea to run the free scan first to see if it catches any known vulnerabilities in your setup before committing to a deeper integration.
For anyone pushing MCP into production, robust observability is an inevitable requirement. Spanly is one of the early movers in this space and definitely worth keeping an eye on.
Comments
No comments yet
Be the first to comment