LangChain

Give a LangChain agent Onekey's tools through the hosted MCP server. The MCP adapters turn the eleven tools into LangChain tools; the agent finds endpoints, reads their contracts and fetches data.

Prerequisites#

A Onekey API key in ONEKEY_KEY. For unattended agents, use a key with a credit limit.
A chat model and its API key (any provider LangChain supports).
Python: pip install langchain langchain-mcp-adapters, plus your model provider's package. JavaScript: npm install @langchain/mcp-adapters langchain, plus your provider's package.

The 5-minute path#

The adapters connect over the MCP "http" transport (Streamable HTTP) with custom headers, which is how Onekey authenticates.

import asyncio, os
from langchain.agents import create_agent
from langchain_mcp_adapters.client import MultiServerMCPClient

async def main():
    client = MultiServerMCPClient({
        "onekey": {
            "transport": "http",
            "url": "https://api.onekey.dev/mcp",
            "headers": {"Authorization": f"Bearer {os.environ['ONEKEY_KEY']}"},
        }
    })
    tools = await client.get_tools()
    agent = create_agent("anthropic:claude-sonnet-5", tools)  # or any model string your provider supports
    result = await agent.ainvoke({"messages": "Find the Onekey endpoint for the latest release of a GitHub repo and use it for facebook/react. Say what it cost."})
    print(result["messages"][-1].content)

asyncio.run(main())

The agent will typically run search_endpoints (free) to find github.latest_release, describe_endpoint (free) to read its parameters, then call_endpoint with platform: "github", endpoint: "latest_release" and params: {"owner": "facebook", "repo": "react"}. Each tool result includes credits_used and credits_remaining, so the agent can report the cost.

Example: a dependency check#

Prompt the agent with a list and let it fan out:

What that costs: github.latest_release is 1 credit and github.repo is 1 credit, so 6 credits for three repositories, plus nothing for the discovery calls. If the agent instead uses github.search_repos (5 credits) to find repositories by name, it is 5 credits per search. Tell it the exact endpoints to keep the bill predictable.

To fetch a whole list in one go from your own code rather than through an agent, use batch: it is not exposed as an MCP tool.

Cost notes#

The five free tools (list_platforms, search_endpoints, describe_endpoint, get_status, get_credits) cost nothing; the six tools that spend credits (call_endpoint, lookup_url, search_multi, creator_card, search_universal, report) spend credits.
A cache hit is free and a failed call is refunded.
Prompt for a plan and a price before a job that pages or fans out over platforms. Results carry a pagination.next_cursor when there is more; the agent should page only as far as it needs.
To enrich rows with sentiment or similar signals, see Labels. It is separate from the tools above and is not available through MCP.

Troubleshooting#

Symptom	Likely error	What to do
Every call is refused	`UNAUTHORIZED`	The key is missing, misspelled or revoked. Check the header name and that the value has no stray spaces or quotes.
Calls stop working partway through a job	`INSUFFICIENT_CREDITS` or `KEY_BUDGET_EXCEEDED`	Nothing was charged for the refused call. Top up, or raise the key's credit limit. Do not retry in a loop.
"Too many requests"	`RATE_LIMITED`	Wait the `Retry-After` seconds. Loops that call once per item should add a pause. See Rate limits.
A parameter is rejected	`INVALID_PARAMS`	Read the parameter names from the endpoint's page (or `describe_endpoint`); do not guess them.
A platform fails intermittently	`UPSTREAM_ERROR` / `UPSTREAM_TIMEOUT`	The source platform, not you. Refunded; retry with backoff and check status.
The adapter raises an exception instead of the model seeing an error	any	The adapters raise on tool failure rather than passing it to the model. Wrap the agent call in `try/except` (`try/catch` in JavaScript), and log the `error.type` from the message.
Results end with a `truncated` field	none	The result was shortened to about 30,000 characters. Ask for fewer rows, or page with `cursor` (`INVALID_CURSOR` means the cursor was edited or stale; restart from the first page).

Untested#

The connection shape (transport: "http", url, headers), get_tools / getTools and create_agent / createAgent follow LangChain's current MCP documentation (Python repository, JavaScript guide) as of 2026-09-30. These snippets were not run. Package and function names move quickly, so check the linked pages if an import fails.
The model string anthropic:claude-sonnet-5 is copied from LangChain's page; use the model you have access to.
The hosted MCP server was tested with the official MCP client, not with LangChain's adapters.

Next#

MCP server, Credits & pricing, Errors.