Open main menu

How to Build Your Own Instinct AI Assistant in 2026

Build an Instinct-like AI assistant you text: Agent 37 always-on host from $3.99/mo plus Sendblue iMessage.

Published on

•

Table of Contents

Do not index
This image promotes building an 'Instinct-like AI' using an always-on host combined with an iMessage bridge projected for 2026. It features a smartphone displaying AI usage stats and three key features: 'Text native', 'BYOK' (Bring Your Own Keys), and 'You own keys'. An arrow points from the phone to a cloud server icon, illustrating the AI's cloud hosting. The service 'Agent 37' is offered from $3.99 per month, providing 1 vCPU, 4 GB RAM, and 8 GB storage, utilizing OpenClaw and Sendblue technologies. The image emphasizes personal control and secure key management for AI deployment.
You cannot self-host Instinct. Here's what I'd build instead.
Instinct (instinct.com) is an invite-only personal assistant from Spear Street Technology. You text or call it. It runs on a persistent cloud computer with connected apps. No new app UI. What you can ship today is the same product shape: an always-on agent computer plus an iMessage (or SMS) bridge, with keys and isolation you control. On Agent 37 that stack starts at $3.99/mo (1 vCPU, 4 GB RAM, 8 GB storage) for OpenClaw, Hermes agent, Claude Code, Codex, OpenCode, or Grok CLI, then you wire Sendblue for blue-bubble messaging.

Why Instinct is going viral

Instinct blew up for a few concrete reasons, not because another chatbot launched with a nicer landing page.
First, access is invite-only. Scarcity does the marketing. People share screenshots of waitlist emails the same way they shared Clubhouse codes. Second, the positioning is sharp: no new interfaces. You text it. You call it. That alone feels different when every other product wants you to install an app, learn a sidebar, and open a dashboard.
Third, the demos that travel are life admin, not coding tricks. Follow-ups. Rides. Bookings. Handyman scheduling. A personal assistant that lives in iMessage feels like magic next to "yet another AI chat tab," especially if you're a founder who already lives in Messages all day.

If you're on the waitlist (or already in)

Maybe you signed up for the waitlist weeks ago and heard nothing. Or maybe you got the invite and want to know what's under the hood.
Either way, this guide is for you. If you don't have an invite, you're the main audience. You don't have to wait. You can build the same ops shape today: messaging front door, always-on computer, tools, your keys. If you do have Instinct, you'll still recognize the architecture below. I'm not cloning their proprietary model. I'm rebuilding the harness around open tooling you can own.

What is Instinct AI?

Instinct is a personal AI assistant you reach by text or call. Official positioning: there are no new interfaces. It connects to applications and devices (email, messaging, screen, audio, location, and more), uses a phone and a computer the way a human would, and handles everyday life tasks such as follow-ups, rides, bookings, and handyman scheduling. Access is private / waitlist-based while they scale compute.
Independent writeups describe the same architecture idea: a messaging front door, a persistent cloud machine, stored credentials for tools, and proactive follow-up. That's the pattern I rebuild in this guide with open tooling. Background reading: Vellum's Instinct breakdown and Instinct's own about page.
One honesty check before we go further: a DIY stack will not clone Instinct's proprietary model or private invite product. It clones the ops shape: always on, text-native, tool-using, your data boundaries.

Instinct-like architecture: three layers

This image illustrates an 'Instinct-like architecture' stack comprised of three main components: Sendblue, Agent 37, and OpenClaw or Hermes or Claude Code. Sendblue is responsible for iMessage API and webhooks; Agent 37 represents an isolated always-on instance; and OpenClaw or Hermes or Claude Code serves as the BYOK (Bring Your Own Knowledge) brain. The architecture is designed for messaging combined with always-on computing and runtime, and emphasizes that no Mac Mini is required for blue bubbles. The image branding at bottom right indicates 'Agent 37.' This architecture is intended for systems that require continuous runtime and messaging integration with flexible brain modules.
If you're evaluating agent harnesses for an ambitious product, start here. The stack is three layers. Everything else is implementation detail.
Layer
Instinct (managed product)
Your build
Interface
iMessage / WhatsApp / call
Sendblue iMessage/SMS API + webhooks
Always-on computer
Spear Street cloud machine
Agent 37 isolated instance from $3.99/mo
Agent runtime
Proprietary Instinct model + tools
OpenClaw, Hermes agent, or Claude Code (+ BYOK)
App connections
Connected accounts on their host
Composio / MCP / skills you wire yourself
Ownership
Their ToS and infra
Your keys, whitelist, and per-instance isolation
Sendblue is the messaging pipe. Docs: Sendblue getting started, AI agents use case, webhooks, MCP. Agent 37 is the machine that stays awake so the bot answers when your laptop sleeps.

Which runtime should you pick for an Instinct-like bot?

Before you pick a runtime, ask yourself what the bot is actually for. Life admin over text is a different job from a coding assistant you message from your phone.
Runtime
Best when
iMessage path
OpenClaw
Personal assistant with browsers, dashboard, broad channels
openclaw-sendblue plugin
Hermes agent
Memory continuity and lean overnight ops
Sendblue webhook or community plugin (verify live docs)
Claude Code
Repo-aware coding assistant you text
Sendblue MCP + always-on host. Session must stay open.
Most "Instinct-like life admin" builds I'd start on OpenClaw, because browsers and tool breadth match travel, bookings, and logged-in sites. Hermes wins if continuity is the product. Claude Code wins if the job is code, not life admin. Compare: Hermes vs OpenClaw vs Claude Code.

How to build it on Agent 37 + Sendblue (step by step)

Here's the path I'd actually take. No Mac Mini. No waiting on Spear Street.
This image outlines a five-step process for setting up Agent 37 with Sendblue, guiding from an empty account to achieving the first successful communication (blue bubble). The steps include launching an Agent 37 instance starting at $3.99 per month, setting up Sendblue with key management, installing the openclaw-sendblue plugin and configuring whitelisting via allowFrom, adding tools such as Composio or using a browser on Pro plans, and finally securing privacy through isolation and review gates. It serves as a concise instructional overview for users integrating these tools.

1. Launch an always-on agent computer

You need a machine that does not sleep when you close your laptop.
  1. Sign up at agent37.com.
  1. Pick Basic ($3.99/mo) for a lean pilot, or Pro ($29.99/mo) when you need Chromium for browser-heavy OpenClaw or Claude Code work.
  1. Launch an isolated OpenClaw (recommended default) or Hermes / Claude Code instance.
  1. Open the browser terminal. Add BYOK model keys.
Same dashboard also hosts Codex, OpenCode, and Grok CLI if your Instinct clone grows into a multi-runtime fleet. Deploy OpenClaw walkthrough: how do you deploy OpenClaw.

2. Provision Sendblue for iMessage

This is the blue-bubble pipe. You are not standing up Apple hardware.
  1. Install the CLI: npm install -g @sendblue/cli
  1. Run sendblue setup (or phone setup per docs).
  1. Run sendblue show-keys and save API Key, API Secret, and your Sendblue number.
  1. Whitelist / verify contacts before production. Free shared-line plans require verified recipients. Dedicated / AI Agent plans remove that restriction (confirm live pricing on sendblue.com).
Architecture reminder from Sendblue: inbound iMessage hits your webhook, your agent thinks, then you POST a reply through the Sendblue API. No Mac Mini required for the blue-bubble pipe.

3. Connect the channel to your runtime

Now glue the pipe to the brain. Pick one path. Don't mix them on day one.
Path A: OpenClaw + openclaw-sendblue (best Instinct-like default)
Community plugin documented in Sendblue's getting started:
  1. Clone and build openclaw-sendblue into your OpenClaw extensions folder on the Agent 37 instance.
  1. Add apiKey, apiSecret, phoneNumber, and allowFrom (your personal number) to OpenClaw config.
  1. Restart the gateway. Text the Sendblue number from an allowed phone.
Path B: Claude Code + Sendblue MCP
  1. Keep Claude Code always on (Agent 37 Basic or Pro). Laptop sleep kills channel bridges.
  1. Add the Sendblue MCP server per MCP docs so Claude can send/receive and manage contacts from the session.
  1. Optionally use community daemons such as TextMe if you want a dedicated Claude-over-iMessage loop. Still host the process on Agent 37, not a sleeping Mac.
Path C: Custom webhook agent
  1. Expose a small HTTPS webhook on your instance (or a tiny relay) that accepts Sendblue receive events.
  1. Acknowledge with HTTP 200 fast, then run your LLM / tools.
  1. Reply with Sendblue send-message. Deduplicate webhook deliveries.
Webhook types and management: Sendblue webhooks.

4. Add tools so it can act like Instinct

Instinct's magic is not the chat bubble. It is the computer that can use apps. On Agent 37:
  • Use Composio (and similar) for Gmail, calendar, Notion, Slack, and 1000+ apps where your runtime supports it.
  • Prefer Pro when the agent must drive a real browser.
  • Store secrets in the instance env, not in screenshots or public git.
  • Start with a short allowlist of tasks (inbox triage, calendar holds, research briefs) before you grant shopping or booking permissions.

5. Lock down privacy before you invite friends

If you're shipping this for yourself, fine. If you're evaluating it as product infrastructure, do not skip this section.
Consumer Instinct raised public debate about broad data licenses and always-logged-in credentials. Your clone should be explicit:
  • Whitelist allowFrom numbers. Do not leave the line open to the world.
  • One Agent 37 instance per person or per client. Do not share multitenant boxes.
  • Separate BYOK keys. Rotate when staff changes.
  • Document what tools can spend money or send email without a human confirm.
  • Agencies: white-label API from $1.99 per agent per month, billed per minute, when you productize this for clients.

Instinct vs your Agent 37 build

This image provides a side-by-side comparison of two similar software product builds: Instinct and Agent 37 combined with Sendblue. It highlights their key differences such as availability—Instinct being on an invite waitlist versus Agent 37 ready to ship today, deployment options—Instinct using their cloud computer compared to Agent 37 offering isolated instances starting at $3.99 per month, technology usage—proprietary model for Instinct and 'Bring Your Own Key' (BYOK) support for Agent 37, and communication methods—text or call for Instinct versus iMessage via API for Agent 37. The visual aims to inform potential users or stakeholders about product ownership distinctions and feature differences.
Here's the honest side-by-side. Same shape. Different ownership.
Dimension
Instinct
DIY on Agent 37 + Sendblue
Access
Invite / waitlist
You ship today
Interface
Text / call
iMessage via Sendblue (+ other channels on OpenClaw/Hermes)
Computer
Their cloud
Your isolated Agent 37 instance
Model
Their proprietary stack
BYOK (Claude, GPT, open models)
Ops
They own it
Platform patches on Agent 37; you own prompts and tools
Entry host cost
Not publicly self-serve
From $3.99/mo + Sendblue line + tokens
At $50/hour, one afternoon of Mac Mini babysitting exceeds a year of Agent 37 Basic ($47.88). Instinct-like bots fail when the machine sleeps, not when the prompt is mediocre.

What it costs

Don't compare "free software" to Instinct. Compare the full bill: host, line, tokens.
Line
Typical shape
Agent 37 Basic
$3.99/mo · 1 vCPU / 4 GB / 8 GB
Agent 37 Plus
$9.99/mo · 1 vCPU / 4 GB / 12 GB
Agent 37 Pro
$29.99/mo · 2 vCPU / 8 GB / 20 GB · Chromium
Agent 37 Max
$99.99/mo · 4 vCPU / 16 GB / 40 GB
Sendblue
Free sandbox / shared line for tests. Production lines billed by Sendblue (verify live AI Agent / dedicated plans)
Model tokens
BYOK. Often the largest line once the bot is chatty

Frequently Asked Questions

Can I literally self-host Instinct?

No. Instinct is a private-access product on Spear Street Technology infrastructure. This guide builds an Instinct-like assistant you own.

Do I need a Mac Mini for iMessage?

Not for the Sendblue path. Sendblue provides the blue-bubble API and webhooks. Your Agent 37 instance is the always-on brain.

Is OpenClaw required?

No, but it is the closest default for browser-heavy life admin. Hermes fits memory-first ops. Claude Code fits coding assistants you text.

Is this safe for client work?

Only with whitelist, isolation, review gates, and clear tool permissions. Do not paste production secrets into community Discord screenshots.

Where do I start today?

  1. Agent 37 Basic at $3.99/mo, OpenClaw template.
  1. Sendblue setup + sendblue show-keys.
  1. Install openclaw-sendblue, whitelist your number, send a test text.
  1. Add one tool (calendar or email) before you add ten.