Introducing Cloud MCP Router: Progressive Tool Discovery for Real‑World AI Agents

TL;DR — Agents don’t fail because they can’t call tools; they fail because we dump all the tools into the prompt at once. Cloud MCP Router adds a progressive, discovery‑driven layer in front of your MCP servers so the model sees only what it needs, when it needs it. Smaller prompts, fewer mis‑selections, and the ability to scale past arbitrary tool caps—without rewriting your existing MCP servers.

The problem we’re solving

Modern agents juggle calendars, email, docs, CRMs, repos, issues, and PRs. Hand the model 80–200 tools at once and you get:

Context overload: Tool schemas crowd out user content and inflate token cost.
Decision paralysis: Long, look‑alike tool lists cause mis‑selections and retries.
Arbitrary caps: Teams hide tools just to keep prompts short—shrinking what agents can actually do.

This pattern is especially acute in Model Context Protocol (MCP) ecosystems designed for rich, tool‑heavy workflows. The result: your “power user” assistant never reaches its potential because the tool layer isn’t designed for discovery.

Meet Cloud MCP Router

Cloud MCP Router is a progressive, discovery‑driven layer that sits in front of your existing MCP servers (official, custom, or community) and turns any large toolset into a query‑as‑you‑go catalog. Instead of spraying every tool and schema into the base prompt, it exposes a small set of router tools that guide the agent through staged narrowing—only surfacing schemas at the moment of execution.

What the model sees

Start with discovery: discover_server_actions returns just the relevant actions (by intent), not entire schemas.
Drill in on demand: get_action_details reveals parameters only for the chosen action.
Execute confidently: execute_action runs with the now‑known parameters (with elicitation automatically bridged).
Stay unblocked: search_documentation fetches the smallest useful doc snippets.

How it works (under the hood)

1) Collision‑safe proxying of any MCP server

Point the Router at remote MCP servers (stdio) and it will connect, list their tools/prompts/resources, and proxy them through a clean namespace. It sanitizes and uniquifies names, enforces length limits, and pre‑checks for collisions before exposing anything to the client—so your model never sees confusing duplicates.

Why that matters: Tools stay stable and human‑readable, even across many vendors and teams. No more “mystery collisions” that silently break calls.

2) Router‑only mode to shrink the prompt on demand

Flip router‑only mode and the Router removes all proxied tools from the surface area, leaving only the small, discovery‑first router tools. When you’re ready, you can selectively re‑enable a subset of actions or whole servers—and still avoid overload. Toggle via the Router-Mode switch on your dashboard.

3) Progressive discovery & execution API (the “router tools”)

discover_server_actions — intent → relevant actions (fast, ranked).
get_action_details — reveal schema only for the chosen action.
execute_action — call the remote tool; elicitation is automatically bridged back to the upstream client session so the agent can ask follow‑ups mid‑execution without losing context.

4) Resources & prompts, too

Beyond tools, the Router proxies resources (with stable proxy://… URIs) and prompts, keeping names safe and avoiding collisions across servers. It even installs a catch‑all resource template so ad‑hoc reads “just work.”

5) Auth that fits your stack

The Router protects its MCP endpoint with OAuth2 authentication against your Cloud MCP endpoint. It also publishes well‑known protected resource metadata so compliant clients know how to authorize. You get modern headers and correct WWW-Authenticate challenges out of the box.

Why customers adopt Cloud MCP Router

Smaller prompts, lower cost: Only a tiny subset of metadata is surfaced at any moment.
Fewer mistakes: Staged, ranked choices beat a flat list of 100+ tools.
Scale past arbitrary caps: Because schemas are fetched lazily, catalogs can grow to hundreds of actions without overwhelming the model.
Keep your stack: No rewrites required—just add your existing MCP servers and go.
Built‑in resilience: Doc search and auth‑recovery tools unstick agents without inflating the base prompt.
Operational sanity: Health checks, safe namespacing, and collision detection remove brittle edges before they reach production.

Example: What the agent’s loop looks like

Discover: “Open a PR that closes issue #123” → Router returns relevant actions like repos.create_pull_request, issues.update, ranked by intent.
Detail: Model asks details for repos.create_pull_request; Router returns just that action’s schema.
Execute: Model calls execute_action with the filled parameters; if extra info is needed mid‑call, elicitation is bridged back to the client session seamlessly.

Net effect: The base prompt stays lean. The agent makes better choices. Your catalog can grow without fear.

Security & enterprise readiness

Session‑verified access tokens (via Cloud MCP).
Correct WWW-Authenticate challenges and .well‑known metadata for resource authorization.

Ready to route?

Cloud MCP Router is available now. If you’re building assistants that need to span dozens of apps without cratering reliability or cost, this is the missing layer.

Try it with your existing MCP servers—no rewrites.
Book a demo to see progressive discovery reduce errors on your own workflows.
Talk to us about rollout, SSO, and enterprise controls.

Let’s help your agents do more—with less.