@gram-ai/functions

Gram Functions for TypeScript

Gram Functions are small pieces of code that represent LLM tools. They are deployed to Gram and are then exposed to LLMs via MCP servers.

This library provides a small framework for authoring Gram Functions in TypeScript. The "Hello, World!" example is:

import { Gram } from "@gram-ai/functions";
import * as z from "zod/mini";

const gram = new Gram().tool({
  name: "greet",
  description: "Greet someone special",
  inputSchema: { name: z.string() },
  async execute(ctx, input) {
    return ctx.json({ message: `Hello, ${input.name}!` });
  },
});

export default gram;

Quickstart

You can use one of the following command to scaffold a new Gram Function project quickly:

pnpm create @gram-ai/function@latest --template gram

## Or one of the following:
# bun create @gram-ai/function@latest --template gram
# npm create @gram-ai/function@latest -- --template gram

Installation

Use one of the following commands to add the package to your project:

pnpm add @gram-ai/functions

## Or one of the following:
# bun add @gram-ai/functions
# npm add @gram-ai/functions

Core Concepts

The Gram Instance

The Gram class is the main entry point for defining tools. You create an instance and chain .tool() calls to register multiple tools:

import { Gram } from "@gram-ai/functions";
import * as z from "zod/mini";

const gram = new Gram()
  .tool({
    name: "add",
    description: "Add two numbers",
    inputSchema: { a: z.number(), b: z.number() },
    async execute(ctx, input) {
      return ctx.json({ sum: input.a + input.b });
    },
  })
  .tool({
    name: "multiply",
    description: "Multiply two numbers",
    inputSchema: { a: z.number(), b: z.number() },
    async execute(ctx, input) {
      return ctx.json({ product: input.a * input.b });
    },
  });

export default gram;

Tool Definition

Each tool requires:

• name: A unique identifier for the tool
• description (optional): Human-readable description of what the tool does
• inputSchema: A Zod schema object defining the expected input parameters
• execute: An async function that implements the tool logic

Tool Context

The execute function receives a ctx (context) object with helper methods:

ctx.json(data)

Returns a JSON response:

async execute(ctx, input) {
  return ctx.json({ result: "success", value: 42 });
}

ctx.text(data)

Returns a plain text response:

async execute(ctx, input) {
  return ctx.text("Operation completed successfully");
}

ctx.html(data)

Returns an HTML response:

async execute(ctx, input) {
  return ctx.html("<h1>Hello, World!</h1>");
}

ctx.fail(data, options?)

Throws an error response (never returns):

async execute(ctx, input) {
  if (!input.value) {
    ctx.fail({ error: "value is required" }, { status: 400 });
  }
  // ...
}

ctx.signal

An AbortSignal for handling cancellation:

async execute(ctx, input) {
  const response = await fetch(input.url, { signal: ctx.signal });
  return ctx.json(await response.json());
}

ctx.env

Access to parsed environment variables defined by the Gram instance:

const gram = new Gram({
  envSchema: {
    BASE_URL: z.string().transform((url) => new URL(url)),
  },
}).tool({
  name: "api_call",
  inputSchema: { endpoint: z.string() },
  async execute(ctx, input) {
    const baseURL = ctx.env.BASE_URL;
    // Use baseURL...
  },
});

Input Validation

Input schemas are defined using Zod:

import { Gram } from "@gram-ai/functions";
import * as z from "zod/mini";

const gram = new Gram().tool({
  name: "create_user",
  inputSchema: {
    email: z.string().check(z.email()),
    age: z.number().check(z.min(18)),
    name: z.optional(z.string()),
  },
  async execute(ctx, input) {
    // input is fully typed based on the schema
    return ctx.json({ userId: "123" });
  },
});

Lax Mode

By default, the framework strictly validates input. You can enable lax mode to allow unvalidated input to pass through:

const gram = new Gram({ lax: true });

Environment Variables

Defining Variables

Environment variables that are used by tools must be defined when instantiating the Gram class. This is done using a Zod v4 object schema:

import { Gram } from "@gram-ai/functions";
import * as z from "zod/mini";

const gram = new Gram({
  envSchema: {
    API_KEY: z.string().describe("API key for external service"),
    BASE_URL: z.string().check(z.url()).describe("Base URL for API requests"),
  },
});

Whenever a tool wants to access a new environment variable, a definition must be added to the envSchema if one does not exist. When this Gram Function is deployed, end users will then be able to provide values for these variables when installing the corresponding MCP servers.

Runtime Environment

Environment variables are read from process.env by default, but you can override them when creating the Gram instance. This can be useful for testing or local development. Example:

import { Gram } from "@gram-ai/functions";
import * as z from "zod/mini";

const gram = new Gram({
  env: {
    API_KEY: "secret-key",
    BASE_URL: "https://api.example.com",
  },
  envSchema: {
    API_KEY: z.string().describe("API key for external service"),
    BASE_URL: z.string().check(z.url()).describe("Base URL for API requests"),
  },
});

If not provided, the framework falls back to process.env.

Response Types

The framework supports multiple response types. All response methods return Web API Response objects.

JSON Response

return ctx.json({
  status: "success",
  data: { id: 123, name: "Example" },
});

Text Response

return ctx.text("Plain text response");

HTML Response

return ctx.html(`
  <!DOCTYPE html>
  <html>
    <body><h1>Hello</h1></body>
  </html>
`);

Custom Response

You can also return a plain Response object:

return new Response(data, {
  status: 200,
  headers: {
    "Content-Type": "application/xml",
    "X-Custom-Header": "value",
  },
});

Error Handling

Using ctx.fail()

Use ctx.fail() to throw error responses:

async execute(ctx, input) {
  if (!input.userId) {
    ctx.fail(
      { error: "userId is required" },
      { status: 400 }
    );
  }

  const user = await fetchUser(input.userId);
  if (!user) {
    ctx.fail(
      { error: "User not found" },
      { status: 404 }
    );
  }

  return ctx.json({ user });
}

Errors automatically include a stack trace in the response.

Using assert()

The assert function provides a convenient way to validate conditions and throw error responses:

import { assert } from "@gram-ai/functions";

async execute(ctx, input) {
  assert(input.userId, { error: "userId is required" }, { status: 400 });

  const user = await fetchUser(input.userId);
  assert(user, { error: "User not found" }, { status: 404 });

  return ctx.json({ user });
}

The assert function throws a Response object when the condition is false. The framework catches all thrown values, and if any happen to be a Response instance, they will be returned to the client.

Key points about assert:

• First parameter is the condition to check
• Second parameter is the error data (must include an error field)
• Third parameter is optional and can specify the status code (defaults to 500)
• Automatically includes a stack trace in the response
• Uses TypeScript's assertion type to narrow types when the assertion passes

Manifest Generation

Generate a manifest of all registered tools:

import { Gram } from "@gram-ai/functions";

const gram = new Gram()
  .tool({
    /* ... */
  })
  .tool({
    /* ... */
  });

const manifest = g.manifest();
// {
//   version: "0.0.0",
//   tools: [
//     {
//       name: "tool1",
//       description: "...",
//       inputSchema: "...", // JSON Schema string
//       variables: { ... }
//     },
//     ...
//   ]
// }

Handling Tool Calls

Exporting the Gram instance from your module as the default export will allow Gram to handle tool calls automatically when deployed:

import { Gram } from "@gram-ai/functions";

const gram = new Gram()
  .tool({
    /* ... */
  })
  .tool({
    /* ... */
  });

export default gram;

You can also call tools programmatically:

const response = await gram.handleToolCall({
  name: "add",
  input: { a: 5, b: 3 },
});

const data = await response.json();
console.log(data); // { sum: 8 }

With abort signal support:

const signal = AbortSignal.timeout(5000);

const response = await gram.handleToolCall(
  { name: "longRunning", input: {} },
  { signal },
);

Type Safety

The framework provides full TypeScript type inference:

import { Gram } from "@gram-ai/functions";
import * as z from "zod/mini";

const gram = new Gram().tool({
  name: "greet",
  inputSchema: { name: z.string() },
  async execute(ctx, input) {
    // input.name is typed as string
    return ctx.json({ message: `Hello, ${input.name}` });
  },
});

// Type-safe tool calls
const response = await g.handleToolCall({
  name: "greet", // Only "greet" is valid
  input: { name: "World" }, // input is typed correctly
});

// Response type is inferred
const data = await response.json(); // { message: string }