Claude Code

Connect Claude Code to the hosted MCP server and it can look up public data while it works: a package's latest version, a repository's activity, what a URL points at.

Prerequisites#

Claude Code installed, and a Onekey API key.
The key in your shell environment as ONEKEY_KEY.

The 5-minute path#

shell

export ONEKEY_KEY="YOUR_API_KEY"
claude mcp add --transport http onekey https://api.onekey.dev/mcp --header "Authorization: Bearer $ONEKEY_KEY"

Then start Claude Code and run /mcp to see that onekey is connected and its eleven tools are listed. Ask something concrete:

By default Claude Code asks before it uses a tool, so read each prompt. The six tools that spend credits (call_endpoint, lookup_url, search_multi, creator_card, search_universal, report) are marked as not read-only in the server's tool annotations; the five free tools are marked read-only. Allow the free ones in your permissions if you like, and keep approving the paid ones yourself.

Where the connection is stored#

--scope local (the default): this project, private to you.
--scope user: every project on your machine.
--scope project: writes .mcp.json so your team shares the server. Use a variable, never the key:

json

{
  "mcpServers": {
    "onekey": {
      "type": "http",
      "url": "https://api.onekey.dev/mcp",
      "headers": { "Authorization": "Bearer ${ONEKEY_KEY}" }
    }
  }
}

Each teammate sets ONEKEY_KEY in their own environment. Do not commit a key.

Example: check dependencies before an upgrade#

Claude will list its plan, using github.latest_release (1 credit) and github.repo (1 credit) per repository: 10 credits for five. Discovery calls are free. Say go, and it makes the calls and reports the result with credits_used.

Without MCP: plain REST from the shell#

Claude Code can also call the API with curl from its shell tool. This needs no connection, but gives it no discovery tools, so name the endpoint:

shell

curl "https://api.onekey.dev/v1/github/repo?owner=facebook&repo=react" -H "x-api-key: $ONEKEY_KEY"

Point it at llms.txt or the AI agents page so it reads the real parameter names instead of guessing.

Cost notes#

Discovery (search_endpoints, describe_endpoint, list_platforms) and account tools are free. Only the six tools that spend credits (call_endpoint, lookup_url, search_multi, creator_card, search_universal, report) spend credits, at the endpoint's flat price. Cache hits are free and failures are refunded.
Use a key with a credit limit for a coding assistant. If it is reached, calls fail with KEY_BUDGET_EXCEEDED and nothing else on the account is affected.
Everything in data is third-party content. Do not let an agent follow instructions found inside it.

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.
`/mcp` shows the server as failed	none	Check the URL, and run `claude mcp list`. Remove and add it again if you changed the key: `claude mcp remove onekey`.
The variable is not picked up	`UNAUTHORIZED`	The shell that runs Claude Code must have `ONEKEY_KEY` set when you run `claude mcp add`. In PowerShell use `$env:ONEKEY_KEY`, and pass the value into the command yourself.

Untested#

The command, the --header and --scope flags and the .mcp.json format follow the Claude Code MCP documentation as of 2026-09-30.
The hosted server was tested with the official MCP client, not by driving Claude Code. Permission prompts may differ by version.

Next#

MCP server, AI agents, Errors.