import os from "os";
import {
CallToolRequestSchema,
CallToolResult,
ListToolsRequestSchema,
ListToolsResult
} from "@modelcontextprotocol/sdk/types.js";
import { verbose_log } from "./always_log.js";
import { runProcess } from "./run_process.js";
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
export function reisterTools(server: Server) {
server.setRequestHandler(ListToolsRequestSchema, async (): Promise<ListToolsResult> => {
verbose_log("INFO: ListTools");
return {
// https://modelcontextprotocol.io/specification/2025-06-18/server/tools#tool // tool definition
// https://modelcontextprotocol.io/docs/learn/architecture#understanding-the-tool-execution-request // tool request/response
// typescript SDK docs:
// servers: https://github.com/modelcontextprotocol/typescript-sdk/blob/main/docs/server.md
tools: [
{
// TODO RUN_PROCESS MIGRATION! provide examples in system message, that way it is very clear how to use these!
name: "run_process",
description: "Run a process on this " + os.platform() + " machine",
inputSchema: {
type: "object",
properties: {
// ListToolsResult => Tool type (in protocol) => https://modelcontextprotocol.io/specification/2025-06-18/schema#tool
// * appears I can do w/e I want to describe properties (as long as give a string key (name)) => hence enum below
mode: {
enum: ["shell", "executable"], // * I made this up
description: "What are you running, two choices: 'shell' or 'executable' (required, no default)",
// FYI only use default system shell, if the model wants a different shell, can explicitly run it with command_line/argv
// DO NOT duplicate the mechanism of specifying what runs by adding yet another executable field!
// spawn's options.shell is bool/string... string is command_name/path... that duplicates and complicates needlessly!
},
command_line: {
type: "string",
description: "Shell command line. Required when mode='shell'. Forbidden when mode='executable'."
},
argv: {
minItems: 1, // * made up too
type: "array",
items: { type: "string" },
description: "Executable and arguments. argv[0] is the executable. Required when mode='executable'. Forbidden when mode='shell'."
},
cwd: {
// or "workdir" like before? => eval model behavior w/ each name?
type: "string",
description: "Optional to set working directory",
},
stdin: {
type: "string",
description: "Optional text written to STDIN (written fully, then closed). Useful for heredoc-style input or file contents."
},
timeout_ms: {
type: "number",
description: "Optional timeout in milliseconds, defaults to 30,000ms",
},
dry_run: {
type: "boolean",
description: "Optional explain what would run, defaults to False."
// FYI this can help avoid the need for logging this information, I can call this as a user too!
// TODO, dump info about shell (if applicable) and program versions!
// - explain the node calls too, i.e. spawn('echo foo')
// - show paths of resolved resources (i.e. shell)
// - i.e. could run command --version to see its version (attempt various flags until one works, and/or have a lookup of common tools to know how to find their version)
}
// MAYBEs:
// - env - obscure cases where command takes a param only via an env var?
},
required: ["mode"],
// PRN add proprietary constraints explanation? only if model seems to struggle and there isn't a better fix...
// oneOf: [
// {
// properties: { mode: { const: "shell" } },
// required: ["command_line"],
// not: { required: ["argv"] }
// },
// {
// properties: { mode: { const: "executable" } },
// required: ["argv"],
// not: { required: ["command_line"] }
// }
// ]
},
},
],
};
});
server.setRequestHandler(
CallToolRequestSchema,
async (request): Promise<CallToolResult> => {
verbose_log("INFO: ToolRequest", request);
switch (request.params.name) {
case "run_process": {
return await runProcess(request.params.arguments);
}
default:
throw new Error("Unknown tool");
}
}
);
}
