Where the money actually goes
MCP servers are cheap to run and expensive to own. Compute is a rounding error; the real costs are the gateway build (auth, limits, logs), the on-call surface, and the upstream API load that agents generate. Pricing and protection are therefore the same conversation.
Hosted pricing models, compared honestly
- Per endpoint (predictable). One price per published MCP with an included request allowance and overage. Easy to budget; scales with your catalog, not your traffic spikes. This is Command+K's model: free tier (1 endpoint, 2,500 req/mo), then $9.99/endpoint/mo with 10k included, $0.50 per extra 100k.
- Pure usage. Pay per call. Fair at tiny scale, anxiety-inducing at agent scale — one retry loop is a bill.
- Platform bundle. MCP hosting inside a bigger AI platform fee. Fine if you use the platform; expensive as an MCP-only choice.
The protection stack that makes pricing safe
- Rate limit + burst per MCP: absorbs spikes, protects your upstream API. Rejections get HTTP 429 with Retry-After — the signal well-behaved agents obey.
- Monthly cap with notify-or-pause: your stop-loss above the included quota. Auto-resume at the next period means no 3am manual unpausing.
- Per-customer allowances: one customer's runaway loop blocks only that customer — raising the allowance restores them instantly.
- Honest counters: tool calls count; handshakes and auth probes don't. Diagnostics should never eat quota.
Start at zero, keep the stop-loss
The free tier exists to prove the loop with a real client before you pay anything; the caps exist so growth never turns into a surprise invoice. See what's included →
FAQ
- How much does it cost to host an MCP server?
- Self-hosting costs engineering time more than compute: the gateway (auth, quotas, logs) is the real build. Hosted options price per endpoint — Command+K is free for 1 endpoint with 2,500 requests/month, then $9.99 per endpoint/month with 10,000 requests included and $0.50 per additional 100k requests.
- What protects me from a runaway agent driving up usage?
- Layered limits: a per-minute rate limit with burst absorbs spikes; a monthly cap (notify or auto-pause) is your stop-loss; per-customer allowances stop one customer's runaway loop without affecting others. Agents retry aggressively — 429 with Retry-After is the language they understand.
- Should MCP usage count handshakes or just tool calls?
- Meter what maps to value and cost: tool calls. Protocol chatter (initialize, tools/list, pings) and system events (auth probes, rejections) should be visible in logs but excluded from billing counters — otherwise diagnostics inflate customers' bills.