Vercel AI SDK

• Self-hosted Node (your own VM, ECS, Fly.io, Railway, a Vercel Sandbox, etc.): you can mount SMFS as a real filesystem on the server.
• Vercel Functions / serverless / edge: there’s no long-lived process to hold a FUSE mount, so use the Bash Tool (@supermemory/bash) instead. The container becomes the filesystem; no mount needed.

​

How it works

​

Self-hosted Node (real mount)

​

Vercel Functions / serverless (Bash Tool)

​

Prerequisites

• A Supermemory API key
• An Anthropic API key
• For Pattern A only: SMFS installed on your server (curl -fsSL https://smfs.ai/install | bash)

​

Pattern A: Claude Agent SDK on self-hosted Node

smfs login --key $SUPERMEMORY_API_KEY
smfs mount my_agent --path ./memory

import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

MEMORY = "./memory"

async def main():
    async for message in query(
        prompt=f"You have a persistent memory filesystem at {MEMORY}. "
               "Read profile.md to learn about the user, then create "
               "session_notes.md summarizing what you found.",
        options=ClaudeAgentOptions(
            allowed_tools=["Bash", "Read", "Write"],
            cwd=MEMORY,
        ),
    ):
        print(message)

asyncio.run(main())

python3 agent.py

​

Pattern B: Vercel AI SDK + Bash Tool (serverless-friendly)

npm install @supermemory/bash ai @ai-sdk/anthropic zod

import { generateText, tool } from "ai";
import { anthropic } from "@ai-sdk/anthropic";
import { createBash } from "@supermemory/bash";
import { z } from "zod";

export async function POST(req: Request) {
  const { prompt } = await req.json();

  const { bash, toolDescription } = await createBash({
    apiKey: process.env.SUPERMEMORY_API_KEY!,
    containerTag: "my_agent",
  });

  const result = await generateText({
    model: anthropic("claude-sonnet-4-5"),
    tools: {
      bash: tool({
        description: toolDescription,
        inputSchema: z.object({ cmd: z.string() }),
        execute: async ({ cmd }) => bash.exec(cmd),
      }),
    },
    maxSteps: 10,
    prompt,
  });

  return Response.json({ text: result.text });
}

• maxSteps: 10 lets the agent chain multiple bash calls per request (read profile.md, then cat a few notes, then write a summary). Bump it if your agent needs deeper chains; lower it to cap cost per request.
• toolDescription is a pre-written description of the available bash surface (semantic sgrep, cat, ls, redirects, etc.). Hand it straight to the model — don’t roll your own.
• No timeout/abort plumbing. bash.exec already runs against the container over HTTPS, so it returns when the command returns. No event-loop blocking and no FUSE.

​

Tips

• Pattern A: mount SMFS once when your server starts, not per-request. Use --ephemeral if you don’t need a local cache on the server.
• Pattern B: configure memory paths once at startup with configureMemoryPaths(["/notes/", "/journal.md"]) to control which files get distilled into Supermemory memories.
• Both: use smfs grep 'query' (Pattern A) or sgrep 'query' inside the bash tool (Pattern B) for semantic search across all files.