or-info

CLI + MCP server to query OpenRouter model info: prices, ELO rankings, context and comparisons.

Any person or AI agent (Claude Code, Codex, Cursor, pi, etc.) can install it and use it to make informed decisions about which model to use.

Install

npm install -g @aggc/or-info
or-info --version

The npm package is published as @aggc/or-info, but the installed executable is or-info.

You can also run it without a global install:

npx -y @aggc/or-info models --limit 5

Requires Node.js 22 or later.

Install from Smithery

Each release is also published as a Smithery MCPB bundle, importable in one click from clients that support the MCPB format (Claude Desktop, etc.):

https://smithery.ai/server/aggc/or-info

The bundle is attached as an asset to every GitHub release (or-info.mcpb) and can also be installed manually by dropping the file into the client.

Supported runtimes and platforms:

• Node.js 22+
• macOS, Linux, Windows

Config and cache paths

or-info resolves config and cache natively per platform:

| Platform | Config directory | Cache directory | |----------|------------------|-----------------| | macOS / Linux | $XDG_CONFIG_HOME/or-info or ~/.config/or-info | $XDG_CACHE_HOME/or-info or ~/.cache/or-info | | Windows | %APPDATA%\\or-info | %LOCALAPPDATA%\\or-info | | Any platform | OR_INFO_CONFIG_DIR override | OR_INFO_CACHE_DIR override |

Files of interest:

• Config file: <config-dir>/.env
• Model cache: <cache-dir>/models.json
• LMArena cache: <cache-dir>/benchmarks.json

API key (optional)

Without an API key the CLI works with OpenRouter's public catalog. With a key you also see private/pay-gated models.

Bash / Zsh

export OPENROUTER_API_KEY=sk-or-...

PowerShell

$env:OPENROUTER_API_KEY = "sk-or-..."

CMD

set OPENROUTER_API_KEY=sk-or-...

Config file

macOS / Linux:

mkdir -p ~/.config/or-info
echo 'OPENROUTER_API_KEY=sk-or-...' >> ~/.config/or-info/.env

Windows PowerShell:

New-Item -ItemType Directory -Force "$env:APPDATA\or-info" | Out-Null
Add-Content -Path "$env:APPDATA\or-info\.env" -Value "OPENROUTER_API_KEY=sk-or-..."

Windows CMD:

if not exist "%APPDATA%\or-info" mkdir "%APPDATA%\or-info"
echo OPENROUTER_API_KEY=sk-or-...>> "%APPDATA%\or-info\.env"

For tests and debugging you can redirect storage without touching your real machine state:

OR_INFO_CONFIG_DIR=/tmp/or-info-config OR_INFO_CACHE_DIR=/tmp/or-info-cache or-info status

CLI usage

List models

or-info models                        # All models sorted by name
or-info models --sort price           # Cheapest output first
or-info models --sort context         # Largest context first
or-info models --filter coding        # Models whose ID/name contains "coding"
or-info models --free                 # Free models only
or-info models --limit 20             # Limit the number of results
or-info models --tags                 # Show feature tags (vision, tools, reasoning…)
or-info models --json                 # Raw JSON

Pricing and details

or-info price anthropic/claude-sonnet-4.5
or-info price google/gemini-2.5-flash --json

ELO ranking (LMArena)

or-info benchmark openai/gpt-4o
or-info benchmark deepseek/deepseek-r1 --json

Shows the model's ELO score from LMArena (human preference votes), confidence interval, global rank and vote count. Data is fetched live from HuggingFace and cached locally for 24 hours — no API key required.

Compare two models

or-info compare anthropic/claude-sonnet-4.5 google/gemini-2.5-flash
or-info compare openai/gpt-4o deepseek/deepseek-chat-v3-0324 --json

Top models for a task

or-info top --task coding                          # Best coding models
or-info top --task reasoning                       # Best reasoning models
or-info top --task general                         # Best all-rounders
or-info top --task vision                          # Best vision models (requires image input)
or-info top --task cheap                           # Best value for money
or-info top --task premium                         # Highest quality, ignoring price
or-info top --task coding --pricing premium        # Best coder regardless of price
or-info top --task coding --budget 2               # Best coders under $2/M output
or-info top --task general --limit 10

Ranking combines LMArena ELO with price and context window size.

--task controls which ELO category and capability filter to apply. --pricing overrides the price-penalty strategy independently:

| --pricing | Effect | |-------------|--------| | standard (default) | Moderate penalty for expensive models | | cheap | Steep penalty; strongly favours free/low-cost models | | premium | No penalty; ranks by quality alone |

Task defaults (when --pricing is not set):

| Task | Default pricing | Capability filter | |------|----------------|-------------------| | general | standard | none | | coding | standard | soft penalty (−15%) if no tool support | | reasoning | standard | none | | vision | standard | hard filter: image input required | | cheap | cheap | none | | premium | premium | none |

Cache management

or-info status          # Show cache age and TTL for each data source
or-info refresh         # Force-refresh OpenRouter catalog + LMArena ELO

MCP server

or-info can run as an MCP server for AI agents.

Tools available

| Tool | Description | |------|-------------| | models.get | Pricing, context, architecture, features and LMArena ELO for a model | | models.list | List models with optional filter, sort and limit | | models.compare | Side-by-side comparison of two models | | models.top | Ranked top models for coding/reasoning/general/vision/cheap/premium; accepts optional pricing override | | benchmarks.get | LMArena ELO score, global rank, vote count and confidence interval for a model | | cache.refresh | Force-refresh OpenRouter catalog + LMArena ELO |

Legacy flat names (get_model_info, list_models, get_benchmarks, compare_models, best_for_task, refresh_cache) are still advertised in tools/list as deprecated aliases (same schemas, prefixed [Deprecated]) and remain callable. The dot-notation names are the canonical ones.

Register in Claude Code

The recommended way is the claude mcp add command, which writes to ~/.claude.json:

# Global — available in all projects
claude mcp add --scope user or-info -- or-info --mcp

# Project-only (run from the project directory)
claude mcp add or-info -- or-info --mcp

Then verify:

claude mcp list

Project .mcp.json — commit this file to share the config with your team:

macOS / Linux:

{
  "mcpServers": {
    "or-info": {
      "command": "or-info",
      "args": ["--mcp"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "or-info": {
      "command": "or-info.cmd",
      "args": ["--mcp"]
    }
  }
}

Without a global install (macOS/Linux):

{
  "mcpServers": {
    "or-info": {
      "command": "npx",
      "args": ["-y", "@aggc/or-info", "--mcp"]
    }
  }
}

Without a global install (Windows):

{
  "mcpServers": {
    "or-info": {
      "command": "npx.cmd",
      "args": ["-y", "@aggc/or-info", "--mcp"]
    }
  }
}

Register in Codex

Add to ~/.codex/config.toml:

[mcp_servers.or-info]
command = "or-info"
args = ["--mcp"]

Then restart Codex for the change to take effect.

Use from Pi

Pi does not use an mcpServers settings schema. The recommended integration is a Pi skill that calls the installed or-info CLI, for example ~/.pi/agent/skills/or-info/SKILL.md.

Test the MCP server

macOS / Linux:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | or-info --mcp

Windows PowerShell:

'{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | or-info.cmd --mcp

Windows CMD:

echo {"jsonrpc":"2.0","id":1,"method":"tools/list"} | or-info.cmd --mcp

Data sources

| Data | Source | Refresh | |------|--------|---------| | Model catalog and pricing | OpenRouter API | Every 30 min | | ELO rankings | LMArena via HuggingFace dataset | Every 24 h |

ELO data is fetched directly from the lmarena-ai/leaderboard-dataset dataset on HuggingFace using their public Datasets Server API — no API key required. Coverage: ~350 models including all major commercial and open-source models tracked by LMArena.

Testing

npm test

npm test runs the deterministic local suite and is the release gate used before publishing. Live integration tests are available separately because they depend on OpenRouter and HuggingFace availability and can occasionally hit third-party rate limits.

Additional entry points:

• npm run test:local for deterministic no-network coverage
• npm run test:online:smoke for the live smoke subset used by CI as a non-blocking signal
• npm run test:online for the full live CLI/MCP suite, including edge cases

Release

See CHANGELOG.md for release history and details.

Contributing

See CONTRIBUTING.md — adding new CLI commands or new MCP tools.

License

MIT