MCP pricing comparison

If you're evaluating tools to observe MCP cost, the decision matrix is which layer they hook into. Most existing observability platforms instrument the LLM call — they miss MCP tool calls entirely. Here's the honest comparison.

What each tool actually observes

ToolSees LLM callsSees MCP tool callsSees per-tool costFree tierPaid entry
MCPSpend✓ (via the tool envelope)✓ every call✓ native25K calls/mo$29 / mo Pro
Helicone✓ excellentonly LLM-level10K requests/mo$25 / mo
Langfuse✓ tracesonly if you instrument every serveronly what you push50K events/mo$29 / mo
PostHogvia wrappersonly if you push eventsonly what you wire1M events/mo$0 then volume
Portkey✓ gatewayonly if proxied through gatewayLLM-level10K requests/mo$49 / mo
Lunary✓ tracespartialLLM-level1K runs/mo$20 / mo

The crux: where is your money going?

In an agentic workload, the LLM completion is roughly 15-25% of total token cost. The other 75-85% is MCP tool calls. If your observability tool only sees the LLM layer, you're looking at the small slice and missing the elephant.

Helicone, Lunary, Langfuse, Portkey were all built before MCP existed. They're excellent at what they do — chat completion observability — but tool-call attribution isn't their model.

PostHog can technically track anything, but you're building the MCP schema from scratch — there's no out-of-the-box concept of "tool", "server", "cost per call". That's weeks of work and on-going maintenance.

When to pick which

Pick MCPSpend if: you run agentic workflows (Cursor Composer, Claude Desktop with MCP servers, Windsurf, background agents, OpenAI Agents SDK with MCP). You want a single dashboard showing which server / tool / project is burning your tokens, with budgets and alerts. Free 25K/month is usually enough.

Pick Helicone if: your workload is mostly direct LLM chat completion (no MCP), and you want best-in-class LLM-layer dashboards, prompt caching analytics, and routing. Their MCP-layer is not a current focus.

Pick Langfuse if: you ship LLM features at scale and need traces + evals + a place to run offline tests. You can push MCP events from your servers manually but it's extra work.

Pick PostHog if: you already run PostHog for product analytics and want one less vendor. Accept that you'll wire the MCP schema yourself.

Combine: MCPSpend + Helicone is a common pairing — MCPSpend on the MCP layer, Helicone on the LLM layer. You see both ends of the agent loop.

Detailed per-vendor comparisons

We maintain dedicated comparison pages with deeper breakdowns:

Track your own MCP spend — free

One command wraps every MCP client on your machine. 25,000 tool calls/month on the free tier. No card.

Related guides