Know what every
MCP call costs you.
A transparent proxy that wraps your MCP servers and records every tool call: latency, errors, payload size, and per-model cost estimation. Find which tools, teams, and customers are burning your AI budget — before the invoice lands.
- ✓ Works with Claude, OpenAI, and any MCP-compatible client
- ✓ 5-minute install, zero SDK changes
- ✓ EU-hosted infrastructure, encrypted at rest
Preview of MCPSpend dashboard. Example data shown.
MCP tools are eating budgets nobody is tracking.
Every team building with agents is hitting the same wall: tool calls explode in volume, costs spike unpredictably, and there's no clean way to see what's happening until the invoice lands.
Your AI costs are a black box
Vendor bills arrive a month late, mixed across teams, projects, and customers. You react instead of plan.
No attribution, no accountability
You can't answer "which customer cost us $4k last week" or "which workflow is bleeding tokens." Engineering ships blind.
No audit trail when you need one
When finance, security, or a customer asks "what did this AI actually do last Tuesday at 3pm?" — you have no answer. Logs are scattered, lossy, or missing.
Five minutes from install to first insight.
MCPSpend sits transparently between your agents and your MCP tools. No SDK changes, no proxy wars, no vendor lock-in.
Install the proxy
One npm install. Drop @mcpspend/proxy in front of your MCP server config. Five lines of code, zero changes to your agent.
npx @mcpspend/proxy add --key sk_live_...
Route calls through MCPSpend
Every MCP tool call is intercepted, timed, and logged with full context — user, project, customer, latency, tokens.
→ tool.call(args) ✓ logged in 8ms
See exactly where your spend goes
Live dashboards by team, customer, tool, and workflow. Budget alerts. Audit logs. Export to your warehouse.
dashboard.mcpspend.com → real-time
Everything finance, security, and engineering ask for.
Real-time tool-call observability
Every MCP call captured in <100ms. Volume, latency, success rate per server and per tool — live, no batch jobs.
Per-team & per-customer attribution
Scope API keys to projects, tag sessions with team or customer. Slice usage any way finance needs.
Cost estimation per model
Each call is priced from payload tokens × your model's per-token rate. Approximate but defensible — and improving as we add IDE-layer integrations.
SLA & latency monitoring
p50/p95/p99 latency per MCP tool. Know which providers are slow before your customers complain.
Complete audit logs
Every call logged with full provenance — who, when, what, how much. Immutable, queryable, exportable to your warehouse.
IDE extensions (coming soon)
Native VS Code extension Q3 2026, then Cursor and Windsurf. Direct LLM-token attribution per MCP tool, without an external proxy.
Pay for what you ship, not for seats.
Volume-based pricing. No per-user tax. Free tier you can actually run a project on.
Free
For solo builders shipping their first agent.
Start free- ✓25,000 tool calls / month
- ✓1 project, 7-day retention
- ✓Real-time dashboard
- ✓Community support
Pro
For small teams running agents in production.
- ✓1M tool calls / month
- ✓Unlimited projects
- ✓30-day retention
- ✓Budget alerts & Slack/email
- ✓CSV + S3 export
- ✓Email support
Team
For multi-team orgs with attribution needs.
- ✓10M tool calls / month
- ✓Per-team & per-customer attribution
- ✓90-day retention
- ✓SAML SSO
- ✓Priority support
- ✓BigQuery / Snowflake export
Enterprise
For regulated industries and large deployments.
- ✓Unlimited calls
- ✓DPA (Data Processing Agreement)
- ✓Dedicated infra (single-tenant) available
- ✓Audit logs (unlimited retention)
- ✓Custom SLA
- ✓Dedicated support contact
All paid plans: payment first → email magic link to set your password → instant dashboard access. Billed by NewRzs SRL (CUI RO48756557) · processed securely by Stripe. Cancel anytime, keep your data.
Explore by topic
Specific cost-tracking guides per IDE, model provider, and team use case.
MCP cost — the complete guide
What an MCP tool call actually costs and how to control it.
Read →Cursor MCP cost tracking
Why Cursor users see Anthropic bills jump 3-5× without warning — and what to do.
Read →Claude Desktop cost tracking
Per-server, per-tool spend visibility for Claude Desktop power users.
Read →ChatGPT MCP cost tracking
Tracking MCP tool spend when your agent uses OpenAI models.
Read →MCP pricing comparison
MCPSpend vs. Helicone, Langfuse, PostHog — what each one tracks (and misses).
Read →AI expense tracker for MCP
Per-customer, per-project AI cost attribution at the tool-call layer.
Read →MCP for budgeting
Set monthly budgets per agent, get alerts before you overrun.
Read →MCP for procurement
What procurement teams need to know before signing AI agent contracts.
Read →Enterprise MCP pricing
Volume pricing, SLA, DPA, audit log retention for larger teams.
Read →Honest answers to honest questions.
- How does MCPSpend calculate cost per tool call? +
- Our proxy sees the MCP wire — request and response payloads on stdio. We approximate input/output tokens from JSON size (≈ chars/4) and multiply by your declared model's per-token rate. It's a useful estimate, not a billing invoice. Direct LLM-layer token tracking is coming via IDE extensions.
- What about LLM tokens from Cursor / Claude Code / Copilot subscriptions? +
- Those tokens live inside the IDE's LLM call, not the MCP wire — so our proxy can't see them directly. We're building VS Code (then Cursor and Windsurf) extensions that hook into the IDE's AI layer to report exact token counts per MCP call. Q3 2026 timeline.
- Does the proxy add latency to my MCP calls? +
- The proxy is fire-and-forget. It forwards bytes between your agent and MCP server immediately, then logs metadata to MCPSpend asynchronously. If our API is unreachable, your agent still works — events buffer in memory and drop after 1000 if we can't reach the server.
- Do you store the tool inputs or outputs themselves? +
- No. We log only metadata: tool name, server name, model, latency, success/error code, payload size in tokens. The actual arguments (filenames, search queries, document contents) never leave your machine.
- Can I export the data? +
- Yes. Pro+ tiers include CSV export; Team+ includes BigQuery and Snowflake destinations. We don't lock you in — your data is yours.
- What runtime do I need? +
- Node.js 18+. No install required — the proxy is published as @mcpspend/proxy on npm and runs via npx. One command (`npx @mcpspend/proxy init --key …`) detects every supported MCP client on your machine and wraps the servers for you.
Stop guessing what AI costs you.
One command to install. Free tier with 25,000 tool calls/month, no credit card.