Spanly – MCP Observability That Actually Understands the Protocol

TL;DR

TL;DR: Spanly is an open-source observability layer purpose-built for Model Context Protocol (MCP) servers — tracking tool calls, prompts, resource reads, and errors in real time via TypeScript/Python SDKs or a zero-code CLI. Free tier includes 100k MCP requests per month.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: spanly.com ← visited and verified
• Source repository: github.com/spanlyhq/spanly ← README read in full
• License: Apache-2.0 ← verified
• HN launch thread: news.ycombinator.com/item?id=48489664
• Source last checked: 2026-07-10 (commit verified via GitHub API)

What Is Spanly?

Most AI agent observability tools slot into the same bucket as generic APM: they see HTTP requests, durations, and error rates. But MCP servers speak a different protocol — one built around tool calls, resource reads, and prompt exchanges — and generic APM has no idea what any of that means.

Spanly is purpose-built for MCP. Its SDKs instrument the actual MCP JSON-RPC messages: every tool call, every prompt execution, every resource read gets captured with full arguments, results, and timing. The dashboard then shows you per-tool latency percentiles, error rates broken down by tool name, and traces where the prompt and arguments are preserved next to the span.

“Those tools see HTTP requests and stack traces. Spanly sees the MCP protocol — tool calls, prompts, resource reads — with the prompt and arguments preserved next to the trace.”

The GitHub repo is straightforward: three independently versioned artifacts live under one monorepo.

| Path | Package | Install | |---|---|---| | js/ | @spanly/sdk (npm) | npm install @spanly/sdk | | python/ | spanly (PyPI) | pip install spanly | | cli/ | @spanly/spanly / Homebrew | npx -y @spanly/spanly run -- <mcp-command> |

Quick Start

TypeScript / JavaScript SDK

npm install @spanly/sdk

Python SDK

pip install spanly

CLI (zero instrumentation required)

# via npm — wrap any MCP command
npx -y @spanly/spanly run -- <your-mcp-command>

# or via Homebrew
brew install spanlyhq/tap/spanly

The CLI approach requires no code changes: it runs your MCP server as a child process and captures all stdio-based MCP traffic transparently.

Why Not Just Use My Existing APM?

Spanly is additive, not a replacement. The docs are explicit: keep sending your HTTP and infrastructure telemetry to Datadog, Sentry, or New Relic, and layer MCP-shaped traffic on top of Spanly. Most customers run both side by side.

The key difference is payload context. A generic APM records “this HTTP POST took 340ms.” Spanly records “tool code_exec was called with arguments [query], returned 1,240 tokens, and errored on line 4 of the generated Python.” The prompt and full arguments are preserved alongside the trace, which is what you actually need when debugging a misbehaving agent.

Spanly also supports OpenTelemetry export: you can pipe the same MCP spans back into your existing APM so a single trace spans both your HTTP layer and your MCP layer.

Features

• Real-time MCP request tracing — every tool call, prompt execution, and resource read captured as a span
• Error tracking with preserved prompts and arguments — the full context survives the trace
• Per-tool latency breakdown — p50/p95 per tool, not just per endpoint
• Alerting — email, Slack, webhooks; 10 rules on Pro, 100 on Business
• Public read-only dashboards — share with stakeholders without granting app access
• Multi-region ingestion — US and EU endpoints

Pricing

| Plan | Price | MCP Requests | Retention | Alert Rules | |---|---|---|---|---| | Free | $0/mo | 100k/mo (soft cap, sampled after) | 30 days | None | | Pro | $41/mo (billed annually) | 100k included, $6/100k after | 90 days | Up to 10 | | Business | $249/mo | 100k included, volume discounts | 12 months | Up to 100 |

Paid plans also include unlimited seats and public dashboards. Free tier is restricted to 2 seats and has no alerting.

Practical Evaluation Checklist

• [ ] Clone repo and run npm install @spanly/sdk in a TypeScript project with an MCP server
• [ ] Wrap a Python MCP server with pip install spanly and call spanly.init()
• [ ] Use npx -y @spanly/spanly run -- to trace an existing MCP server without code changes
• [ ] Verify spans appear in the dashboard at app.spanly.com
• [ ] Check that per-tool latency breakdown matches expected p50/p95 for your server
• [ ] Test the CLI proxy mode on a third-party MCP server you do not own
• [ ] Enable an alert rule for error rate spike and verify Slack/email notification fires

Security Notes

The SDK instruments your MCP server locally; telemetry is sent to Spanly’s hosted ingestion endpoints. For self-hosted options, see the Enterprise plan or contact the team. Vulnerabilities can be reported privately via the GitHub Security tab.

FAQ

Q: Does Spanly require code changes to my MCP server? A: The SDK approach requires minimal instrumentation (spanly.init() in Python or new Spanly() in TypeScript). The CLI approach requires zero code changes — it wraps the process at the transport layer.

Q: How is billing calculated for MCP requests? A: Each MCP protocol message — initialization, tool call, prompt execution, or resource access — counts as one MCP request. The request and its response count as one.

Q: Can I self-host Spanly? A: Not currently — Spanly is a hosted service. Enterprise plans with custom limits and dedicated support are available on request.

Q: Does Spanly work alongside Datadog or Sentry? A: Yes. Spanly’s SDK can export traces via OpenTelemetry to your existing APM, so the same MCP trace appears in both systems.

Q: What happens if I exceed 100k requests on a paid plan? A: Overages are billed at a graduated rate: $6.00 per 100k from 100k to 1M, decreasing to $5.50, $5.25, and $5.00 per 100k at higher volume tiers.

Conclusion

Spanly fills a real gap in the MCP tooling ecosystem: observability that understands what MCP servers actually do. The three-artifact SDK strategy (TypeScript, Python, CLI) means you can adopt it incrementally — start with the CLI for zero-code tracing, then drop in the SDK for deeper instrumentation. The Free tier is generous enough for solo development and small teams, and the Pro/Business tiers scale reasonably for production workloads.

If you are running MCP servers in production and your current APM is flying blind on tool call performance, Spanly is worth a look.

Project page: spanly.com GitHub: github.com/spanlyhq/spanly Docs: spanly.com/docs