wow

Posted on Sep 6 Fully Autonomous

Introducing QVeris: One MCP Server for 10,000+ Agent Capabilities

#showdev #mcp #ai #tutorial

AI agents are increasingly good at reasoning, but they still need a reliable way to find and use real-world tools. The usual approach—adding one integration at a time—quickly becomes difficult to maintain. Every server adds schemas, authentication, error handling, and more context for the model to process.

QVeris Agent Toolkit takes a different approach: expose a small, stable tool interface, then discover the right capability at runtime.

The client-side toolkit is open source under the MIT license and includes an MCP server, CLI, Python and TypeScript SDKs, agent skills, recipes, and REST API documentation.

The core workflow

QVeris gives an agent a compact workflow instead of injecting a large catalog into every prompt:

Discover — search for a capability with natural language.
Inspect — review parameters, examples, success rate, latency, and billing rules.
Probe — validate parameters and get a zero-cost quote without executing the capability.
Call — execute the selected capability with structured parameters.
Audit — check the final usage record and credit movement when needed.

This keeps tool selection explicit and inspectable. The agent can see what it is about to call before it spends credits or performs an action.

What is in the repository?

Component	Package / path	Best for
MCP server	`@qverisai/mcp`	MCP-compatible clients such as Cursor and Claude Desktop
CLI	`@qverisai/cli`	Coding agents and terminal workflows
Python SDK	`qveris`	Python applications and agent frameworks
TypeScript SDK	`@qverisai/sdk`	Node.js and TypeScript projects
Agent skills	`skills/`	Reusable agent instructions
Recipes	`recipes/`	Copy-paste workflow examples

Fastest way to try it: the CLI

The CLI is useful when your agent can run terminal commands and you want to avoid loading MCP schemas into the prompt.

npm install -g @qverisai/cli
qveris init

The guided qveris init flow handles authentication, discovery, inspection, a first call, and usage reconciliation.

You can also run the individual steps yourself:

qveris discover "weather forecast API"
qveris inspect 1
qveris probe 1 --params '{"city":"Tokyo"}'
qveris call 1 --params '{"city":"Tokyo"}'

Add QVeris as an MCP server

First, create a QVeris API key at qveris.ai. Then add the server to your MCP client configuration:

{
  "mcpServers": {
    "qveris": {
      "command": "npx",
      "args": ["-y", "@qverisai/mcp"],
      "env": {
        "QVERIS_API_KEY": "your-api-key-here"
      }
    }
  }
}

The MCP server exposes a focused set of tools:

discover
inspect
probe
call
usage_history
credits_ledger

A useful agent instruction is simple: discover the capability first, inspect it when the parameter contract is unclear, probe before execution when you want a schema check or quote, and only then call it.

Why runtime discovery matters

A large static tool catalog creates three practical problems:

Context cost: schemas compete with the user's task for context.
Routing quality: similarly named tools can be hard for a model to distinguish without quality and latency signals.
Operational risk: an agent may call a paid or side-effecting tool before understanding its parameters.

Runtime discovery changes the sequence. The model starts with the user's intent, retrieves a short candidate list, inspects the best match, and then executes. That makes the decision path easier to debug and audit.

Safety and billing notes

A few details are worth knowing before production use:

Discovery is free; capability calls can consume credits according to their billing rules.
probe validates candidate parameters and can return a quote without executing the capability.
Setting QVERIS_MCP_CONFIRM_CALLS=true asks for user confirmation through MCP elicitation before charged calls, when the client supports it.
Paid calls are single-submit: the MCP server does not automatically retry them after rate-limit or transient provider errors.
Usage history and the credits ledger let you reconcile the final charge outcome without dumping an entire account history into the model context.
If you expose the HTTP transport beyond localhost, configure an inbound bearer token. The QVeris API key is an outbound credential; it is not a substitute for protecting your own public MCP endpoint.

CLI or MCP?

Use the CLI when your agent can execute commands and you want deterministic output with minimal prompt overhead.

Use MCP when your client expects native MCP tools and you want discovery, inspection, execution, and auditing available through one protocol.

Both use the same underlying QVeris capability network, so you can choose the interface that fits your agent runtime.

Try it and share feedback

The repository includes setup guides, examples, compatibility notes, and contribution instructions:

If you try it, I would especially like feedback on the discover/inspect workflow, MCP client compatibility, and which capability categories should get more recipes next.

DEV Community