Connect LangGraph to Skysay

Give your LangGraph application access to Skysay's phone-number, call and SMS capabilities through hosted MCP. Start by proving one authenticated read; then add only the tools your workflow needs.

This first graph performs an account read, not a phone call. It has no model dependency or model charges. Your existing model can plan a task later; Skysay's configured voice AI agent or your audio integration speaks on a call. LangGraph orchestration is not itself a voice engine.

Before you start

• Python 3.13 and uv.
• A Skysay workspace and an API key with account:read.
• Keep the key in a secret manager or the environment that launches Python. Never paste it into a chat, committed file or command-line argument.

Create a scoped key. The first connection uses https://api.skysay.ai/mcp/read-only; its credentials are scoped to this resource. Moving to the full MCP resource is a separate permission/configuration change, not an edit to make casually.

Run the public example

View the example directory. Save these raw files in the same directory:

• requirements.lock
• verify_connection.py
• README.md

In the directory containing those files, run:

uv venv --python 3.13.13
uv pip sync --require-hashes requirements.lock
uv run --no-sync python verify_connection.py

Set SKYSAY_API_KEY securely before the last command. The sample pins langgraph==1.2.15 and langchain-mcp-adapters==0.3.2; the universal lock also pins transitive dependencies. These are tested versions, not a promise that all future releases behave identically.

Expected output:

Verified: LangGraph completed a scoped Skysay account read. No call or SMS sent.

The graph loads the read-only server's tools, selects exactly account_overview, invokes it and ends. Other discovered tools are not executed. It does not print account details or retain them in graph state. The entrypoint prints a generic failure if authentication, discovery or tool execution fails.

How the graph is wired

from langgraph.graph import END, START, StateGraph

builder = StateGraph(State)
builder.add_node("verify_connection", verify_connection)
builder.add_edge(START, "verify_connection")
builder.add_edge("verify_connection", END)
graph = builder.compile()

State and verify_connection are defined in the downloadable verify_connection.py, not built-in functions. Its MCP connection uses:

client = MultiServerMCPClient({
    "skysay": {
        "transport": "streamable_http",
        "url": "https://api.skysay.ai/mcp/read-only",
        "headers": {"Authorization": f"Bearer {key}"},
    }
}, handle_tool_errors=False)

key comes from SKYSAY_API_KEY. The example's allowlist limits its tool invocations; the MCP resource and API-key scopes enforce access on the server. Do not turn on tracing or persist tool outputs without considering what account, message or call data they can contain.

Extend to a telephone workflow

After connection verification, build a separate path with these stages:

1. Collect the requested task, destination, caller number and intended outcome.
2. Check number capabilities, activation, outbound policy and the price basis.
3. Show the exact proposed call or SMS to the person authorizing it.
4. Pause for approval; changing the recipient, message or purpose needs approval again.
5. Execute the approved action once, retain its ID and inspect its result.

A prompt telling a model to ask is not an enforcement boundary. Put the pause in application control flow, outside automatic tool selection. If you use LangGraph interrupts, configure a checkpointer and thread identity appropriate for your application; the minimal graph here does not implement that lifecycle. Do not persist credentials in checkpoint state.

Use the REST first-call guide, policy and outbound requirements, SMS guide and call results for the actual action contracts. A key with account:read cannot place calls. A returned call ID does not mean the person answered or the task completed.

Troubleshooting

The sample deliberately prints a generic failure message; it does not display HTTP status codes or account data. Check your key, its scopes and resource on the workspace API keys page to distinguish the cases below. Do not print the raw exception or tool response to diagnose it.

• 401: check the key, revocation and resource URL. A read-only-resource token does not become a full-resource token by changing the URL.
• 403: verify the key has account:read in the intended workspace.
• Missing tool: use the hosted read-only endpoint and the pinned versions; do not silently switch to another tool or endpoint.
• Graph succeeds but no call happens: expected. This example verifies access only; no paid action node is present.

Verification and sources

The sample is verified against a real local Skysay HTTP handler with an isolated workspace and a disposable scoped key, including authentication failures. No production workspace, hosted model, call or SMS is used. A local pass does not prove your hosted credentials, carrier activation or model behavior.

Official references checked 15 October 2026: LangChain MCP integration, LangGraph Graph API and MCP adapters.