- “How many invoices did we issue last week?”
- “Create a new customer called ACME Inc. with address 123 Main St”
Getting started
Run the server locally with npx:list_api_endpoints
→ get_api_endpoint_schema
→ invoke_api_endpoint
) to fulfil your requests.
Which client should I use?
- Cursor – best for developers working inside a repo/IDE with inline tool usage and project/global config files.
- Claude Desktop – great general‑purpose client on macOS/Windows; easy JSON config and strong tool orchestration.
- Claude (web) – use when you want quick remote access via a URL without local config files.
- ChatGPT (Connectors) – available in certain plans; currently expects
search
/fetch
tools and may need a wrapper for full compatibility.
Client setup
Cursor
Cursor supports MCP via project and global config files and works with stdio or remote HTTP/SSE transports. Config files- Project:
.cursor/mcp.json
- Global:
~/.cursor/mcp.json
.cursor/mcp.json
~/.cursor/mcp.json
npx -y conductor-node-mcp --transport=http --port=3000
. If your Cursor build supports auth headers in mcp.json
, pass an Authorization
bearer token or use the alternative header x-conductor-secret-key
. Otherwise, prefer stdio or a trusted local network. See: Cursor Docs: Model Context Protocol.
Claude Desktop
Local (stdio) configurationclaude_desktop_config.json
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json
- Windows:
%APPDATA%\\Claude\\claude_desktop_config.json
Claude (web)
Claude on the web supports connecting to remote MCP servers. Run Conductor in Remote mode (HTTP) and add it from Claude’s integrations/settings UI, then approve tool calls as needed. See: MCP Clients – modelcontextprotocol.ioChatGPT (Connectors)
ChatGPT supports “Custom Connectors (MCP)” in Settings → Connectors. See: OpenAI Help: Connectors in ChatGPT Compatibility note- ChatGPT may show the error “This MCP server doesn’t implement our specification” if required tools are missing. As of September 2025, Connectors expect servers to implement tools named
search
andfetch
. Our Conductor MCP server exposes explicit tools per endpoint or dynamic tools such aslist_api_endpoints
/invoke_api_endpoint
, so it won’t directly satisfy that requirement without a thin wrapper server that providessearch
/fetch
and delegates to Conductor. See the troubleshooting section in the help article for details.
Advanced options
Remote mode (HTTP)
Run the MCP server as a remote server using Streamable HTTP transport:- Auth via header
Authorization: Bearer <token>
orx-conductor-secret-key: <token>
. - You can also use
--socket
to bind to a Unix socket.
http://localhost:3000?resource=cards&resource=accounts&no_tool=create_cards
http://localhost:3000?client=cursor&capability=tool-name-length%3D40
Choosing the tool style
You have two options when starting the server:- Explicit tools – one MCP tool per Conductor endpoint. Useful when you know exactly which operations you need and want the most accurate parameter suggestions.
- Dynamic tools (
--tools=dynamic
) – three generic tools that let the LLM search, inspect, and invoke any endpoint on demand. Helpful when you want the entire API surface in a compact form.
--tool=<name>
– include a specific tool by name--resource=<pattern>
– include all tools under a resource (supports wildcards, e.g.qbd.invoices*
)--operation=read|write
– include just read or write operations--tag=<tag>
– include tools by tag- Exclusions available via
--no-tool
,--no-resource
, etc. - Use
--list
to preview which tools will be exposed
Capabilities & clients
Different LLM clients have different schema limitations. Pass--client=<type>
so the MCP server tailors its output accordingly. Supported values: openai-agents
, claude
, claude-code
, cursor
.
Fine-tune capabilities with --capability
flags (comma-separated or repeated):
top-level-unions
– enable top-level unions in tool schemasvalid-json
– allow JSON string parsing for argumentsrefs
– enable$ref
pointers in schemasunions
– enableanyOf
union typesformats
– enable format validations likedate-time
tool-name-length=N
– set maximum tool name length