# Kosha Model Discovery [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/sriinnu/kosha-discovery  
**GitHub Stars:** 3  
**npm Downloads (last month):** 3136  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/kosha-model-discovery

## Description
Tells your agent which model to use and what it costs, across 45 providers and local runtimes.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "kosha-model-discovery": {
    "command": "npx",
    "args": ["-y","-p","@sriinnu/kosha-discovery","kosha-mcp"]
  }
}
```

## Documentation

## What Kosha Model Discovery does

Kosha Model Discovery gives MCP-compatible agents access to a registry of AI models, providers, pricing, capabilities, context limits, credentials, and serving routes. It is designed for decisions such as finding the least expensive model that supports tool use and a required context window, or determining which route is available for a model alias.

The package also includes a TypeScript library, CLI, HTTP API, OpenAI-compatible proxy, and spend ledger, but the MCP entry point is `kosha-mcp`. The MCP server runs over standard input and output, so an agent can use it without exposing a separate HTTP service.

## How it works

On startup and during discovery, Kosha checks configured provider access and local credential sources, then combines the results with public catalogs from models.dev and LiteLLM and a curated offline model list. Provider keys are optional. With no credentials, the server can still answer from publicly available model metadata; available credentials allow it to identify models the user can access.

The registry covers 46 providers and local runtimes. Model records include pricing and context information where available, along with capabilities and provider relationships. Alias resolution lets an agent use names such as `sonnet` and obtain the corresponding canonical model ID.

## Setup and configuration

Install the package with Node.js 22 or newer:

```bash
npm install @sriinnu/kosha-discovery
```

Register the stdio server with Claude using:

```bash
claude mcp add kosha -- kosha-mcp
```

Clients that read the MCP registry can install the published server identifier `io.github.sriinnu/kosha-discovery`. Provider credentials are optional, and the server can run on a fresh machine using its public catalogs and offline data.

## Tools and capabilities

The MCP tool set includes:

- `kosha_query_models` for searching the model registry.
- `kosha_cheapest_model` for finding lower-cost models that meet requirements.
- `kosha_ranked_routes` for ranking available serving paths.
- `kosha_model_detail` for inspecting a model record.
- `kosha_model_routes` for listing routes for a model.
- `kosha_resolve_alias` for mapping an alias to a canonical model ID.
- `kosha_provider_health` for checking provider status.
- `kosha_context_strategy` for selecting a model or approach based on context needs.

These tools expose the discovery portion of Kosha to an agent. The separate proxy can resolve aliases, choose providers, forward OpenAI-compatible requests, and record spend, but those HTTP proxy operations are not required for the MCP server.

## Limitations and notes

Discovery quality depends on the catalogs and credentials available to Kosha. Google, Bedrock, and Vertex are listed as discovery-only in the broader package, while some other non-chat provider formats are also not proxied. The MCP server itself uses stdio, so clients must support launching a local MCP process. Provider credentials are not required for basic public-catalog queries, but credential-aware selection depends on the relevant credentials being available to Kosha.

_Full upstream README: https://allmcps.com/mcp/kosha-model-discovery/readme_