salpa-mcp

Let your own AI agent drive your own Salpa, over the Model Context Protocol.

Salpa runs scientific workflows (molecular simulation, docking and similar) on your computer, one step per node on a canvas. With salpa-mcp, an agent you already use can list Salpa's workflow templates, prepare a workflow from one, change its steps' settings, run it, and read what each step made, through the same local API the app's own window uses.

Setting it up

Add it to your agent's MCP configuration as a stdio server, with the command:

uvx salpa-mcp

For example, in a client that reads an mcpServers table:

{
  "mcpServers": {
    "salpa": { "command": "uvx", "args": ["salpa-mcp"] }
  }
}

Start Salpa before you ask your agent to use it. salpa-mcp reaches the Salpa running on this computer, as you, on port 18000 unless SALPA_PORT or SALPA_APP_URL says otherwise. When Salpa is not running, every tool says so.

It needs Python 3.10 or later, which uvx chooses for you, and Salpa 0.12.0 or later. Changing a step's settings (set_node_options) needs Salpa 0.15.0 or later.

The tools

What it will not do

• Change a workflow you have open in Salpa. Salpa's canvas keeps its own copy of a workflow it has open, and would put the change back. Salpa refuses the edit, and the tool tells your agent to ask you to close the workflow's tab, or to make the change in Salpa.
• Reach anything but your own Salpa. It talks only to the app on this computer, with the app's per-launch local API token, which only you can read and which it sends only to the app that issued it.

A run starts work on your computer, and a cloud step (Salpa Compute) uses your account's Salpa Compute allowance. salpa-mcp tells your agent to ask you before it runs anything, and most agents also ask before each tool call.

Licence

Apache-2.0.