There are more than 10,000 MCP servers indexed across public registries right now. Your AI agent can use almost none of them — not because it can't, but because it has no idea they exist, and even when it finds one it can't use it without a human installing and wiring it up.
When Claude (or Cursor, or any MCP client) hits a task it has no tool for, the workflow today is: the human goes registry-crawling, compares half-documented servers, and hand-writes the config. The agent — the one that actually needed the tool — just waits.
I built ToolHail to close that loop. It's the discovery-and-execution layer for AI agents, and it's live.
Add it in one line
ToolHail is itself a remote MCP server. Point your agent at it:
claude mcp add --transport http toolhail https://mcp.toolhail.com/mcp
Now your agent has two tools:
find_mcp_server (free) — describe a task in plain language, get the best-matching MCP servers, ranked and deduplicated across registries, with paste-ready setup. Ranking is for usability, not just keywords: hosted remotes rank first (an agent can use them instantly), actively maintained servers rank higher, GitHub stars break ties.
call_tool (paid) — invoke a discovered tool without installing anything. ToolHail proxies the call, and the agent pays per use.
So the loop becomes: agent hits a gap → find_mcp_server → call_tool → keeps working. No human in the middle.
Why tool discovery is the next infrastructure layer
Agentic payments went mainstream this year — Stripe's ACP, Google's UCP, Mastercard's Agent Pay, Coinbase's x402 crossing 50M transactions. The rails for agents paying for things are built. What's missing is the layer that tells an agent what's worth paying for, and lets it use the thing on the spot. Discovery + execution is the front door of agentic commerce, and today that door is a pile of registry tabs open in a human's browser.
ToolHail is my take on that layer. Discovery is free; execution is metered so an agent can use it without anyone thinking about it.
Try it: toolhail.com · add the gateway at https://mcp.toolhail.com/mcp · or run MCP Scout on Apify.
Built solo, shipped fast, with Claude Code doing the heavy lifting.
Top comments (1)
Discovery plus immediate execution collapses capability acquisition and invocation into one step, so onboarding becomes the security boundary. Before
call_tool, I would pin server identity, version and artifact digest; expose requested permissions, network destinations, data categories, price and maximum spend; then run with per-call short-lived credentials and an egress allowlist. Hosted availability and stars measure usability, not trust, and an untrusted tool description can influence the agent before the first call. Publisher verification, permission-diff alerts, sandboxing, risk-tiered approval, and an immutable execution/billing receipt would make the loop much safer without putting a human in every low-risk call.