PICA MCP Integration with LangChain

PICA provides a unified API platform that connects AI agents to third-party services (CRMs, email, calendars, databases, etc.) through MCP tool calling.

PICA MCP Server

PICA exposes its capabilities through an MCP server distributed as @picahq/mcp. It uses stdio transport — it runs as a local subprocess via npx.

MCP Configuration

{
  "mcpServers": {
    "pica": {
      "command": "npx",
      "args": ["@picahq/mcp"],
      "env": {
        "PICA_SECRET": "your-pica-secret-key"
      }
    }
  }
}

• Package: @picahq/mcp (run via npx, no install needed)
• Auth: PICA_SECRET environment variable (obtain from the PICA dashboard https://app.picaos.com/settings/api-keys)
• Transport: stdio (standard input/output)

Environment Variable

Always store the PICA secret in an environment variable, never hardcode it:

PICA_SECRET=sk_test_...

Add it to .env and load with python-dotenv.

Using PICA with LangChain

LangChain provides MCP client support via the langchain-mcp-adapters package. Always refer to the latest docs before implementing. See langchain-mcp-reference.md.

Required packages

pip install langchain-mcp-adapters langgraph langchain-anthropic mcp python-dotenv

Before implementing: look up the latest docs

The langchain-mcp-adapters API has changed across versions (e.g., MultiServerMCPClient is no longer a context manager as of v0.1.0). Always check the latest docs before writing code. See langchain-mcp-reference.md.

Integration pattern

1. Create an MCP client using MultiServerMCPClient with stdio transport pointed at npx @picahq/mcp
2. Get tools from the client via await client.get_tools()
3. Create a ReAct agent using create_react_agent(model, tools) from langgraph.prebuilt
4. Stream or invoke the agent with your messages
5. Pass environment variables (PICA_SECRET, PATH, HOME) to the MCP client's env config

Minimal example

from langchain_anthropic import ChatAnthropic
from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent

model = ChatAnthropic(model="claude-haiku-4-5-20251001", streaming=True)

client = MultiServerMCPClient({
    "pica": {
        "command": "npx",
        "args": ["@picahq/mcp"],
        "transport": "stdio",
        "env": {
            "PICA_SECRET": os.environ.get("PICA_SECRET", ""),
            "PATH": os.environ.get("PATH", ""),
            "HOME": os.environ.get("HOME", ""),
        },
    },
})
tools = await client.get_tools()
agent = create_react_agent(model, tools)

# Invoke
result = await agent.ainvoke({"messages": [{"role": "user", "content": "..."}]})

# Or stream events
async for event in agent.astream_events({"messages": messages}, version="v2"):
    kind = event["event"]
    if kind == "on_chat_model_stream":
        content = event["data"]["chunk"].content
        # content may be a list of content blocks (Anthropic models) or a string

Important: Anthropic content blocks

When streaming with astream_events(version="v2"), Anthropic models return chunk.content as a list of content blocks, not plain strings. Always handle both:

if isinstance(content, list):
    text = "".join(
        block.get("text", "") if isinstance(block, dict) else str(block)
        for block in content
    )
elif isinstance(content, str):
    text = content

Checklist

When setting up PICA MCP with LangChain:

• langchain-mcp-adapters, langgraph, langchain-anthropic, mcp are installed
• PICA_SECRET is set in .env
• .env is loaded via python-dotenv (load_dotenv() at top of file)
• MCP client uses stdio transport with npx @picahq/mcp
• PATH and HOME are passed in the MCP client env config
• MultiServerMCPClient is NOT used as a context manager (API changed in v0.1.0)
• Streaming handles both list and string content from Anthropic models
• Tool events (on_tool_start, on_tool_end) are handled for UI rendering

Additional resources

• For LangChain MCP adapter docs and API details, see langchain-mcp-reference.md