# legion [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/faulkj/legion-mcp  
**GitHub Stars:** 1  
**npm Downloads (last month):** 483  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/legion

## Description
MCP-native LLM councils for debates, juries, blind panels, refinement, and custom deliberation.

## Tools
Capabilities this server exposes over MCP:

- **poll.md**
- **cancel.md**
- **prompt**
- **quorum**

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

```json
"mcpServers": {
  "legion": {
    "command": "npx",
    "args": ["-y","legion-mcp"]
  }
}
```

## Documentation

## What legion MCP server does

The legion MCP server turns configured language models into MCP tools. Each model file produces a separate tool named from its filename, allowing the calling model to choose among available models based on their descriptions. A model call can include a prompt and optional context, role, system instructions, temperature, and token limits.

The `quorum` tool sends one request to multiple selected models. It supports assigned roles, parallel or multi-round discussion, visibility settings, and synthesis, returning the individual answers separately. Configured presets provide named councils for workflows such as debates, courtroom-style judgments, code review, and other custom arrangements. Each preset is exposed as its own tool.

## How it works

All model requests use the OpenAI Responses API wire format. OpenAI and Azure endpoints can be used directly; other providers need an endpoint or gateway that accepts this format, such as LiteLLM. Legion does not include provider-specific adapters or a built-in model catalog.

Calls are stateless and use `store: false`. The server does not maintain a database or conversation history. Model identity and telemetry are returned in structured content rather than added to the answer text, while logs are written to stderr for stdio safety.

## Setup and configuration

Running the legion MCP server requires Node.js 24 or newer and at least one OpenAI Responses-compatible endpoint. The npm setup can be started with `npx legion-mcp`; a source checkout uses `npm install` and a local `.env` file.

At least one real model file must exist under `config/models/`. npm's bundled example files contain no keys and are ignored by the model scanner, so an npm installation needs a user-provided model file. Each file specifies a deployed model identifier and can optionally define its endpoint, key, baseline system instructions, omitted request parameters, reasoning effort, and maximum output tokens. API keys may be loaded from environment variables with the `env:` prefix.

Configuration also supports Markdown role files, JSON preset files, prompt and error text, schemas, and a description. Local configuration overlays bundled defaults by file or key. HTTP reloads configuration for each request; stdio reads it once per connection, so edits take effect after the client reconnects.

## Tools and capabilities

- Individual tools for every configured model.
- `quorum` for ad-hoc multi-model calls.
- Preset-specific tools for pre-staffed councils.
- Roles with reusable or inline instructions.
- Parallel, multi-round, visibility, and synthesis options for council workflows.
- Per-model limits and parameter omission for endpoint-specific behavior.

The separate `prompt` tool is also exposed by the server, alongside `quorum` and the generated model and preset tools.

## Limitations and notes

The server needs at least one configured model and a compatible endpoint. It does not adapt provider-specific APIs itself, persist conversations, or provide a database. Bundled configuration cannot be deleted through the overlay mechanism; local files can override or add entries. Stdio clients must reconnect before configuration edits become active.

_Full upstream README: https://allmcps.com/mcp/legion/readme_