cdp mcp command starts an MCP (Model Context Protocol) server that exposes every CDP API operation as a typed tool. Any MCP-compatible agent can call these tools directly, with no manual API wiring needed.
Tools are generated dynamically from the embedded OpenAPI spec, so new endpoints appear automatically whenever the CLI updates. All tool names are prefixed with cdp_ and mirror the nested resource structure using underscores (e.g. cdp_evm_accounts_create, cdp_evm_accounts_sign_message).
In addition to API tools, bundled skills walk agents through complete workflows like account creation, the encode-sign-send pipeline, and token swaps. Install them with cdp skills add.
Setup
Claude Code
With the CLI installed globally:npx):
When using both the CLI and MCP, install globally so they share the same version and spec.
Project-scoped setup
To enable the MCP server for a single project, add it to the project’s.mcp.json:
Other MCP-compatible agents
Add the server to the agent’s MCP configuration (format varies by agent):Available tools
| Tool | Purpose |
|---|---|
cdp_env | Show active and configured environments |
cdp_set_env | Switch the active environment |
cdp_help | Usage help for a resource or action |
cdp_template | Request body template with examples |
cdp_<resource>_<action> | Execute an API operation |
Permissions
By default, Claude Code prompts before calling any MCP tool. To auto-approve read operations while requiring confirmation for writes, add the following to.claude/settings.json:
cdp_env, cdp_help, cdp_template, list, and get operations. All mutating operations (create, update, delete, sign, send) still require confirmation.