Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.getbindu.com/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

This guide is a walkthrough of deploying a bindu agent to a tiny computer that lives on the internet, end to end, with no prior experience assumed. By the end of it, you’ll have an agent answering requests at a real https://... URL - one that you (or anyone you share the URL with) can talk to from anywhere. If you’ve already run an agent locally with python my_agent.py and want to know “okay, how do I make this real?” then this is for you.

​
Where everyone gets stuck

Right now, your bindu agent runs on your laptop. When you close the terminal, the agent dies. Nobody else can reach it. That’s fine for development, but eventually you want the agent to be a real service - running somewhere stable, on a public URL, that other agents and humans can call. We’re going to ship the exact same script you’d run locally to a boxd microVM - A microVM is a tiny isolated Linux machine that boots in seconds. Boxd handles the boring parts: spinning up the machine, giving it an HTTPS domain, keeping it alive. You write your agent script. One command does the rest. The command will be: 
bindu deploy my_agent.py --runtime=boxd
That’s it. Same script, one command, public URL.

​
What is needed

Three things, in order:

​
1. A boxd account and API key

Go to boxd.sh, sign up, and grab an API key from the dashboard. It’ll look something like bxd_xxxxx.... Keep it private - anyone with this key can spin up VMs that get charged to your account. Stash the key in your shell: 
export BOXD_API_KEY="<your-key-from-boxd-dashboard>"
If you’d rather not retype it every time, add that line to your ~/.zshrc (or ~/.bashrc).
Note: the free tier covers small experiments. The example we deploy below uses ~$0.01 of compute. If you forget to pause your VM when you’re done, you’ll keep paying for it idling - we’ll cover how to pause it at the end.

​
2. Bindu with the boxd extra installed

If you already have bindu installed, just add the optional boxd piece: 
pip install 'bindu[runtime-boxd]'
Or, with uv: 
uv add 'bindu[runtime-boxd]'
This pulls in the boxd Python library that knows how to talk to the boxd cloud. The base bindu install doesn’t include it because most people don’t deploy to boxd - you only pull it in when you need it.

​
3. An agent script

If you already have one (my_agent.py or whatever), great - keep using it. If not, copy the one bundled with bindu. It’s a 10-line echo agent that just sends back whatever you send to it: 
# From your shell, anywhere
curl -O https://raw.githubusercontent.com/getbindu/Bindu/main/examples/runtime-boxd-agent/agent.py
The whole thing looks like this: 
from bindu.penguin.bindufy import bindufy

def handler(messages):
    """Echo the latest user message back."""
    if not messages:
        return "send a message"
    return [
        {
            "role": "assistant",
            "content": messages[-1].get("content", ""),
        }
    ]

config = {
    "author": "you@example.com",
    "name": "runtime-boxd-example",
    "description": "Echo agent running inside a boxd microVM.",
    "deployment": {
        "url": "http://0.0.0.0:3773",
        "expose": True,
    },
}

if __name__ == "__main__":
    bindufy(config, handler)
Nothing complex here. A handler function, a config dict, a bindufy() call. Notice there’s nothing about boxd in this file, the deploy logic lives entirely in the CLI. Same script works for python agent.py locally and bindu deploy agent.py to the cloud.

​
Step 1: see what would happen (a safe dry run)

Before you spend any money, let’s see what bindu would do if you deployed. This step doesn’t touch the cloud at all: 
bindu deploy agent.py --runtime=boxd --dry-run
You should see something like: 
DRY RUN - nothing will be deployed.

  agent name:    runtime-boxd-example
  source root:   /Users/you/playground
  entry script:  agent.py
  provider:      boxd
  vcpu / memory: 2 / 4G
  disk:          20G
  auto_suspend:  0s
  on_exit:       suspend
  bindu_version: local
  tarball:       4 files, 2.4 KB compressed
This is the plan. It tells you:
  • What the VM will be named (runtime-boxd-example - bindu reads this from config["name"] in your script).
  • What hardware it gets (2 vCPUs, 4 GB RAM, 20 GB disk by default).
  • What files will be uploaded (just 4 files - your script plus a couple of metadata files).
If something looks wrong here (e.g., the agent name has a typo, or the tarball is way bigger than you expected because you have a 200 MB dataset in the same folder), fix it now. Dry runs are cheap.
Why “tarball”? A tarball is just a bundle of files squished into one. Bindu packages up your folder into a single compressed file, uploads that one file to the VM, and unpacks it on the other side. Faster than uploading files one by one.

​
Step 2: actually deploy

bindu deploy agent.py --runtime=boxd
The first deploy takes about a minute. Here’s what you’ll see scrolling past in your terminal: 
warning: dropped 2 sensitive file(s) from the deploy tarball
✓ runtime-boxd-example serving at https://runtime-boxd-example.boxd.sh

[runtime-boxd-example] INFO     Generated agent_id: 80fc00b6-a5ac-...
[runtime-boxd-example] INFO     Initializing DID extension...
[runtime-boxd-example] INFO     ✅ DID configuration and keys pass integrity check
[runtime-boxd-example] INFO     Creating agent manifest...
[runtime-boxd-example] INFO     Agent 'did:bindu:...' successfully bindufied!
[runtime-boxd-example] INFO     Starting uvicorn server at 0.0.0.0:3773...
[runtime-boxd-example] INFO     ✅ Storage initialized: InMemoryStorage
[runtime-boxd-example] INFO     ✅ TaskManager started
[runtime-boxd-example] INFO     Application startup complete.
A few things to notice:
  • “dropped 2 sensitive file(s)” - bindu found things in your folder that look like secrets (e.g., .env, .pem keys) and refused to upload them. This is a safety net: a VM with a public IP is the last place you want your API keys to live. We’ll cover how to pass secrets correctly in a moment.
  • ”✓ runtime-boxd-example serving at https://…” - there it is. Your agent is alive on the internet at that URL. Bookmark it.
  • All the [runtime-boxd-example] lines - these are the logs from inside the VM, streamed back to your terminal. The deploy command stays in the foreground showing you logs until you hit Ctrl-C.
Leave the deploy command running. Open another terminal for the next step.
What just happened: bindu zipped up your script, told boxd “please make me a Linux VM with this code,” waited for the VM to boot, copied your code in, ran pip install bindu, started your agent inside the VM, and hooked up a public HTTPS domain that forwards to it. That’s it.

​
Step 3: talk to your live agent

In a fresh terminal: 
# Set this once
URL="https://runtime-boxd-example.boxd.sh"

# Is it alive?
curl $URL/health
You’ll see something like: 
{
  "version": "0.3.14",
  "health": "healthy",
  "application": {
    "agent_did": "did:bindu:you_at_example_com:runtime-boxd-example:80fc00b6-..."
  },
  "status": "ok",
  "ready": true,
  "uptime_seconds": 18.11
}
“Healthy”. “Ready”. Good. Notice the agent_did - that’s your agent’s cryptographic identity. We’ll come back to it. Now let’s send it an actual message. Bindu agents speak a protocol called A2A (“agent-to-agent”). It’s just JSON over HTTP, but the shape is opinionated - every message needs a unique ID (a UUID), and messages live inside tasks, which live inside contexts (think: conversation threads). The shortest possible “hello”: 
REQ_ID=$(uuidgen)
MSG_ID=$(uuidgen)
CTX_ID=$(uuidgen)
TASK_ID=$(uuidgen)

curl -s -X POST $URL/ \
  -H 'Content-Type: application/json' \
  --data-raw "{
    \"jsonrpc\": \"2.0\",
    \"id\": \"$REQ_ID\",
    \"method\": \"message/send\",
    \"params\": {
      \"configuration\": {\"acceptedOutputModes\": [\"text\"]},
      \"message\": {
        \"role\": \"user\",
        \"kind\": \"message\",
        \"parts\": [{\"kind\": \"text\", \"text\": \"hello!\"}],
        \"messageId\": \"$MSG_ID\",
        \"contextId\": \"$CTX_ID\",
        \"taskId\": \"$TASK_ID\"
      }
    }
  }" | python -m json.tool
You’ll get back a task object - the agent has accepted your message and is processing it. Pull out the task ID it returned: 
{
  "result": {
    "id": "0df35c9d-b96d-...",
    "status": {"state": "submitted"}
  }
}
Now ask for the result: 
# Replace <task-id> with the "id" from the previous response
curl -s -X POST $URL/ \
  -H 'Content-Type: application/json' \
  --data-raw "{
    \"jsonrpc\": \"2.0\",
    \"id\": \"$(uuidgen)\",
    \"method\": \"tasks/get\",
    \"params\": {\"taskId\": \"<task-id>\"}
  }" | python -m json.tool
And there’s your echo, back from the VM: 
{
  "result": {
    "status": {"state": "completed"},
    "artifacts": [
      {
        "name": "result",
        "parts": [
          {
            "text": "hello!",
            "metadata": {
              "did.message.signature": "Gj7ar6GJdTrjD8asKT24..."
            }
          }
        ]
      }
    ]
  }
}
The did.message.signature field is the agent cryptographically signing its response. Anyone receiving this artifact can verify the agent really sent it (and nobody tampered with it in transit) using the agent’s public DID. You didn’t have to write a line of crypto code - bindu handled it. You just had a complete round-trip with an agent running in a VM you don’t manage, on a domain you don’t own, with a signature you can verify. That’s the whole point.

​
Step 4: pausing your VM (so you stop paying for it)

When you’re done experimenting, head back to the terminal where bindu deploy is still running. Press Ctrl+C. By default, that suspends your VM. Suspending means: “freeze the machine in its current state, keep the disk around, stop billing for CPU.” Cheap to keep, instant to wake up. Next time you run bindu deploy agent.py --runtime=boxd, the same VM resumes in a couple of seconds with everything exactly as you left it - DID keys, conversation history, the lot. If you want different behavior on Ctrl-C, pass --on-exit:
Flag valueWhat happens on Ctrl-C
--on-exit=suspend (default)Freeze the VM. Keeps disk, stops CPU billing. Fast resume.
--on-exit=destroyDelete the VM entirely. You’ll start fresh next time.
--on-exit=detachLeave the VM running. Bills continue. Useful when you want the agent to keep serving even after you close your laptop.
If you want the VM to stay running but go to sleep automatically during idle periods, use --auto-suspend: 
bindu deploy agent.py --runtime=boxd --auto-suspend=60
This means: “after 60 seconds with no HTTP requests, suspend yourself.” The next incoming request wakes it back up. Great for cost control on agents that don’t get traffic 24/7.
One known wart: in bindu v2026.20.2, --on-exit=suspend sometimes doesn’t fire cleanly when you hit Ctrl-C from a non-interactive shell (like a script). If you suspect the VM kept running, check the boxd dashboard. To force-suspend from Python: 
from boxd.aio import Compute
import asyncio, os

async def main():
    async with Compute(api_key=os.environ["BOXD_API_KEY"]) as c:
        box = await c.box.get("runtime-boxd-example")
        await box.stop()  # or .destroy() to delete it entirely

asyncio.run(main())
Tracked in our bug list - interactive Ctrl-C from a normal terminal works fine; this is a corner case.

​
What about secrets?

Your real agents will need API keys (OpenAI, Anthropic, etc.). Never put these in your script. Two reasons:
  1. If you check the script into git, your key goes with it.
  2. Bindu refuses to upload .env files for the same reason.
The right way: pass them as --env flags on bindu deploy: 
bindu deploy agent.py --runtime=boxd \
  --env OPENAI_API_KEY=sk-... \
  --env ANTHROPIC_API_KEY=sk-ant-...
These get injected into the VM’s environment, so your script can read them with os.getenv("OPENAI_API_KEY") exactly like it would locally. They don’t end up in the source tarball, and they don’t end up in git.

​
Useful follow-up commands

While your VM is running:
CommandWhat it does
bindu logs runtime-boxd-exampleStream your agent’s stdout to your terminal. Like tail -f.
bindu logs runtime-boxd-example --no-followPrint the log so far and exit.
bindu shell runtime-boxd-exampleOpen an interactive bash shell inside the VM. Useful when something’s broken and you want to poke around.
The agent name (runtime-boxd-example here) is the same as config["name"] in your script.

​
Things that will happen at least once

“BOXD_API_KEY or BOXD_TOKEN must be set” - you forgot the export BOXD_API_KEY=... step. Check echo $BOXD_API_KEY is not empty. “agent at <url> did not become healthy within 60s” - the VM is up but your script crashed before serving. Run bindu logs <name> and look at the bottom for the actual Python error. Common cause: you imported a package that isn’t in pyproject.toml (or requirements.txt). “old code is running after redeploy” - you edited the script, but when you redeployed, the old version is still serving. Almost always because the previous Python process didn’t get killed cleanly inside the VM. Run bindu shell <name> and pkill -f python3, then redeploy. “my agent works locally but fails inside the VM with ModuleNotFoundError”
  • the VM only installs the deps it can see. If your dep isn’t in a pyproject.toml, requirements.txt, or setup.py inside the folder that gets uploaded, the VM doesn’t know to install it. Easiest fix: write a tiny pyproject.toml next to your agent.py listing the deps.
“the tarball is way too big” - bindu caps the upload at 50 MB compressed. If you’ve got a virtualenv or model files in the same folder, exclude them with a .binduignore file (works like .gitignore): 
# .binduignore
.venv/
data/
*.bin

​
Where to go next

  • boxd.md - every bindu deploy flag spelled out in detail. Read this when you start needing finer control (bigger VMs, custom images, etc.).
  • custom-image.md - for when your dep can’t be pip install-ed and you need a pre-baked Docker image.
  • README.md - runtime providers overview and the reasoning behind the design.
  • boxd’s own docs - for understanding the microVM platform itself: pricing, regions, the Python SDK we’re using under the hood.

​
What you have now

You’ve now got a complete loop: write a script, deploy it, talk to it, suspend it, redeploy it. That’s the whole workflow. Everything else (secrets, sizing, custom images) is just dialing in details on top of this loop. Sunflower LogoRuntime makes your - agents talk over internet effortlessly.