Free vs paid AI & ML servers someone always pays for the compute.

First, genuinely free utilities: time-mcp gives a model time awareness with no meaningful marginal cost, and Apollo MCP Server turns your own GraphQL operations into tools — the compute is yours already. Second, free wrappers around paid services: MinerU Open MCP parses PDFs, DOCX, and PPTX into Markdown through the MinerU API, so the server is free but the calls behind it are metered by someone else. Third, free software that runs expensive workloads locally: selvage's LLM-driven code review with AST context extraction is only as free as your model budget.

Before comparing prices, identify which kind you're looking at. The answer determines who is actually paying.

BYO-key works fine for a developer wiring up one assistant. It works badly for autonomous agents: every key is a credential to provision, rotate, and cap, and a leaked key is an open tab on your account. Multiply that across a fleet of agents and a dozen tools — judge servers like mcp-as-a-judge, prompt tooling like Promptheus, each wanting its own backend — and key management becomes its own job.

Rate limits compound the problem. Your key's quota is shared across everything using it, so one chatty agent can starve the rest. None of this shows up on a pricing page, but it's the real cost of the free tier.

An x402-priced endpoint replaces the key with a payment. The agent calls the tool, receives 402 Payment Required with a price, pays in USDC on Base — settlement takes about two seconds — and the request runs. No account, no key to leak, no shared quota; each call carries its own Ed25519-signed receipt, and the operator gets paid before the handler executes. Prices start at $0.01 per call.

For compute-backed tools this maps cost to value precisely. Parsing one contract PDF or reviewing one diff has a knowable marginal cost; per-call pricing lets the operator charge just above it and lets the agent pay only for the calls it makes.

Self-host the free server when the workload runs on compute you already own, the tool wraps your own data (your GraphQL schema, your Anki decks via anki-mcp-server), or you're iterating in development where cost discipline doesn't matter yet. Prefer a paid per-call endpoint when the tool is compute-heavy, the calling agents aren't all yours, or provisioning keys per agent is more operational surface than you want.

Maintainers of AI & ML servers sit on the strongest monetization case in the index: your costs are per-call, so your pricing should be too. Claim your listing, set a price at or above your marginal inference cost, and the gap is margin. Loomal's fee is 5% on settled transactions, currently waived.