MCP Server

The Pyth MCP server gives AI assistants direct, structured access to Pyth market data through the Model Context Protocol. You can discover feeds, fetch real-time and historical prices, and generate chart-ready candles, all within a single conversation.

MCP endpoint: https://mcp.pyth.network/mcp — no self-hosting required.

Tools

Category	Tool	Description
Discovery	`get_symbols`	Find available feeds by query and asset type
Real-time	`get_latest_price`	Fetch latest prices for up to 100 feeds
Historical	`get_historical_price`	Get point-in-time prices for up to 50 feeds
Charting	`get_candlestick_data`	Return OHLC candles (up to 500 per response)

Recommended workflow: get_symbols → get_latest_price / get_historical_price → get_candlestick_data

get_latest_price requires a Pyth Pro access token passed as access_token. Get one at acquire access token guide. All other tools work without a token.

Setup

Endpoint: https://mcp.pyth.network/mcp
Transport: Streamable HTTP
Auth: No key required at the endpoint level

Open Settings

In Claude Desktop, open Settings → Developer → Edit Config.

Add the MCP Server

Add this to claude_desktop_config.json:

{
  "mcpServers": {
    "pyth": {
      "url": "https://mcp.pyth.network/mcp"
    }
  }
}

Restart and Verify

Restart Claude Desktop. MCP tools should appear in chat.

Add Server

claude mcp add pyth --transport http https://mcp.pyth.network/mcp

Verify

claude mcp list

Open MCP Settings

In Cursor, go to Settings → Tools & MCP → Add Custom MCP.

Add Config

{
  "mcpServers": {
    "pyth": {
      "url": "https://mcp.pyth.network/mcp"
    }
  }
}

Verify

Confirm pyth is connected in Settings → Tools & MCP.

Any MCP client that supports Streamable HTTP can connect to:

https://mcp.pyth.network/mcp

Tool Reference

get_symbols

Search and list available Pyth Pro feeds.

Parameter	Type	Required	Description
`query`	string	No	Text search filter
`asset_type`	string	No	`crypto`, `fx`, `equity`, `metal`, `rates`, `commodity`, `funding-rate`
`limit`	number	No	Default `50`, max `200`
`offset`	number	No	Default `0`

get_latest_price

Fetch latest prices for one or more feeds.

Parameter	Type	Required	Description
`access_token`	string	Yes	Pyth Pro token
`symbols`	string[]	No*	Full symbols like `Crypto.BTC/USD`
`price_feed_ids`	number[]	No*	Feed IDs from `get_symbols`
`channel`	string	No	`real_time` or `fixed_rate@{N}ms`
`properties`	string[]	No	Requested fields

* At least one of symbols or price_feed_ids is required.

Max combined feeds per request: 100
If both symbols and price_feed_ids are sent, IDs take precedence
Symbols must include asset-type prefix (Crypto.BTC/USD, not BTC/USD)

get_historical_price

Get prices at a single historical timestamp.

Parameter	Type	Required	Description
`symbols`	string[]	No*	Full symbols
`price_feed_ids`	number[]	No*	Feed IDs
`timestamp`	number	Yes	Unix seconds, milliseconds, or microseconds
`channel`	string	No	`real_time` or `fixed_rate@{N}ms`

* At least one of symbols or price_feed_ids is required.

Max feeds per request: 50
If both IDs and symbols are sent, IDs take precedence
Timestamp is normalized to microseconds and aligned down to channel interval

Historical data is available from April 2025 onward.

get_candlestick_data

Fetch OHLC candlestick data for a single feed.

Parameter	Type	Required	Description
`symbol`	string	Yes	Full symbol (e.g. `Crypto.BTC/USD`)
`resolution`	string	Yes	`1`, `5`, `15`, `30`, `60`, `120`, `240`, `360`, `720`, `D`, `W`, `M`
`from`	number	Yes	Start time (Unix seconds)
`to`	number	Yes	End time (Unix seconds)
`channel`	string	No	`real_time` or `fixed_rate@{N}ms`

from must be earlier than to
Max candles returned: 500
Large result sets are truncated with truncated: true in the response

Troubleshooting

Invalid or expired token

Symptom: get_latest_price returns an auth error
Fix: Verify or regenerate your token and pass it as access_token

Symbol not found

Symptom: Lookup fails for a symbol
Fix: Run get_symbols first and use the exact symbol with its asset-type prefix

No historical data

Symptom: Historical or candle responses are empty
Fix: Use timestamps after April 2025 and ensure the timestamp is not in the future

Truncated candles

Symptom: Response includes truncated: true
Fix: Narrow the time range or use a larger resolution

Local Development

Optionally run the MCP server locally:

git clone https://github.com/pyth-network/pyth-crosschain.git
cd pyth-crosschain
pnpm install
pnpm --filter @pythnetwork/mcp build

Then configure your client with stdio transport:

{
  "mcpServers": {
    "pyth": {
      "command": "node",
      "args": ["<path-to-repo>/apps/mcp/dist/index.js"],
      "env": {
        "PYTH_PRO_ACCESS_TOKEN": "<your-token>"
      }
    }
  }
}

On this page