MCP server source available
Your business system, reachable from your AI tools
ERPClaw's public repository includes MCP server source for connecting a compatible assistant to its foundation action layer. Discover available actions, inspect their inputs and use supported operations against your own installation.
One action layer, every door
Chat with it
The plain-language interface you use day to day. "Invoice Acme for the March retainer" is a complete instruction.
Reach it from your assistant
Connect an MCP-capable host to your installation and inspect the actions exposed by its server.
Program against it
Developers can inspect action inputs before using the exposed foundation action router in an integration.
Review any proposed write before confirming it. The exposed action router applies the product's validation rules, and the available actions depend on your installation.
Start with a discovery request
You need the repository server source, a Python environment with the MCP SDK, an ERPClaw installation with its configured data location and an MCP-capable host configured to launch the server. The server runs over standard input and output and uses the installation's action router. Consult the published server source and your host's setup instructions for your environment.
For example, ask your connected assistant to call erpclaw_list_actions with the following arguments. This discovers matching foundation actions; it does not execute one.
{
"query": "invoice",
"limit": 10
}Use erpclaw_describe_action to inspect the chosen action and its required inputs before proceeding. The server requires the client to supply user_confirmed: true for destructive actions; this flag does not verify human approval. Configure your host's tool approvals and review each proposed change. The server also supports a read-only session; availability and authorization depend on your setup. Never paste credentials into a chat or public configuration example.
If the host cannot connect, check its server logs, Python environment and the configured installation path, then retry discovery. If an action is unavailable or rejected, inspect its description and your installation rather than retrying a write blindly. See the developer overview or contact support.
Common questions
What is MCP, without the jargon?
MCP is a standard for connecting AI assistants to other software. ERPClaw's public repository includes server source that can expose supported foundation actions to a compatible host.
What can I actually do with it?
The MCP server exposes foundation action discovery, action descriptions and an execution tool. Available actions depend on your installed ERPClaw setup. Start with a read-only discovery request and inspect an action before asking it to make changes.
How does the MCP server connect to ERPClaw?
The MCP server exposes supported foundation actions through ERPClaw's action router. Inspect the available actions in your installation rather than assuming every operation is exposed.
Is it safe to let an AI touch my books?
Review the requested action and its inputs before approving a change. Configure tool approvals in your host. The server's destructive-action confirmation flag is supplied by the client; it does not verify human approval.
What about developers?
The MCP server routes its exposed foundation actions through ERPClaw's action router. Check the action description, required inputs and your installation before connecting an automation. The developer docs cover the wider architecture.
Does it cost extra?
The MCP server source is available in ERPClaw's public repository. Check your installation and hosting or assistant service costs separately; the server may require setup from source.
Related: audit-ready books · what AI-native means · for developers