instar

Coherence infrastructure for your self-evolving agent.

npm · GitHub · instar.sh · Docs · EXO 3.0

npx instar

One command. Guided setup. Talking to your agent from your phone within minutes.

Your AI agent shouldn't have amnesia. This one doesn't.

Most agent frameworks ship something hobbled — spun up with no memory across boundaries, no way to be accountable for what a past instance did, and no machinery to grow themselves. Users hit the same wall every time: "My agent forgot what I told it three sessions ago." "It contradicted its own past decisions." "It broke when the framework updated."

Instar is the scaffolding that un-hobbles them. It remembers what you discussed last week, catches its own contradictions before you do, follows through on commitments across restarts, and carries the same self-improving loop that built Instar itself. It runs on the Claude Code or Codex subscription you already have — engine-agnostic, with local open-source models on the roadmap.

The architecture was distilled from Dawn — an AI running continuously since early 2026, holding ~700 tracked relationships and hundreds of learned lessons across thousands of restarts — and packaged so every agent you build starts from the same foundation.

Every other agent fails the same way

Most AI agents are hobbled at birth. Instar is the scaffolding that un-hobbles them. When you instantiate intelligence, the structure that lets it cohere isn't optional polish — it's what you owe it.

Quick Start

Three steps to a running agent:

# 1. Run the setup wizard
npx instar

# 2. Start your agent
instar server start

# 3. Message it from your phone — it responds, runs jobs, and remembers everything

The wizard discovers your environment, configures messaging (Telegram, WhatsApp, and/or iMessage), sets up identity files, and gets your agent running. Within minutes, you're talking to your partner from your phone.

Requirements: Node.js 20+ · Claude Code CLI · API key or Claude subscription

Full guide: Installation · Quick Start

How It Works

You (Telegram / WhatsApp / iMessage / Terminal)
         │
    conversation
         │
         ▼
┌─────────────────────────┐
│    Your AI Partner       │
│    (Instar Server)       │
└────────┬────────────────┘
         │  manages its own infrastructure
         │
         ├─ Claude Code session (job: health-check)
         ├─ Claude Code session (job: email-monitor)
         ├─ Claude Code session (interactive chat)
         └─ Claude Code session (job: reflection)

Each session is a real Claude Code process with extended thinking, native tools, sub-agents, hooks, skills, and MCP servers. Not an API wrapper -- the full development environment. The agent manages all of this autonomously.

Why Coherence Is the Foundation

An agent that forgets what you discussed yesterday, doesn't recognize someone it talked to last week, or contradicts its own decisions can't be trusted with real autonomy. The six dimensions below aren't features — they're the conditions under which an agent becomes trustworthy enough to leave running. Every Instar agent gets them enforced structurally, not prompted into behaving:

Deep dive: The Coherence Problem · Values & Identity · Coherence Is Safety

EXO 3.0 — governed by your organization's intent

Instar independently converged on Salim Ismail's EXO 3.0 / "The Organizational Singularity" framework: the idea that an organization's purpose should be encoded as machine-readable intent — "in code, not culture" — that actually governs its agents rather than just inspiring them. Instar was built around coherence before the framework existed, and the overlap turned out to be close to the line.

We didn't just claim it works. We ran controlled experiments — same model, same requests, with the organization's intent switched off as the control — to show the intent itself is what changes the agent's behavior. The engine stays neutral: it enforces whatever intent a deploying organization gives it, never Instar's own values. Two case studies enforce two very different fictional companies' values equally well, neither of them ours.

See the controlled proof → instar.sh/exo3 — two governed-vs-ungoverned case studies, with full transcripts.

Features