Start Here · 01.01
Welcome to Agentis: The Local-First Operating System for Long-Lived Agents
Agentis is local-first infrastructure for AI agents that keep their memory, identity, and skills for life — even when you swap the model underneath them.
The problem: your agent's mind belongs to someone else
Every agent framework you've used so far makes the same trade, whether it says so or not: the model provider holds the memory. Close the chat window, switch from GPT to Claude, or move from one CLI harness to another, and the agent's accumulated context — what it learned about your codebase, your customers, your preferences — evaporates. You are not building an agent. You are renting one, and the lease resets every session.
Agentis is the alternative: it separates the agent from the model that happens to be executing its current turn. Identity, memory, skills, and reach are durable Agentis state, stored on your machine, versioned like everything else you own. The model is a tenant you can evict and replace without the agent noticing anything changed except its own reasoning got sharper or cheaper.
The mental model: one substrate, six primitives
Agentis is deliberately small at the top. Underneath every feature sits one durable substrate — a restart-survivable state layer — and everything the platform does is one of six primitives, or a composition of them:
| Primitive | Answers the question |
|---|---|
| Agent | Who is doing the work, and what do they remember? |
| Subject | Who is this work being done for — a person or a device? |
| Connection | How does an agent reach a person, a system, or another agent? |
| Orchestration | What is the actual sequence of work, and did it really finish? |
| Experiment | Did this approach actually work better than the alternative? |
| Interface | What does a human (or another program) actually see and touch? |
There is deliberately no seventh primitive. When Agentis grows a new capability, it extends one of these six — it does not bolt on a parallel subsystem with its own rules. That constraint is why the platform stays learnable even as it gets more capable: once you understand the six, you can predict where any new feature lives.
Own the agent, rent the model
This is the ownership stack the whole platform is built around. Models sit at the top, interchangeable and disposable. Everything below the swap seam — the agent's identity, its accumulated Brain, and the durable spine that survives a restart — lives on your machine and belongs to you.
This isn't a slogan bolted onto a normal agent framework after the fact — it's a load-bearing architectural seam called the Runtime Abstraction Layer, and it's the subject of the very next category: Own Your Agents.
What you get, concretely
Embedded SQLite and a bundled offline embedding model. Zero external services required to start. Nothing leaves your machine unless you send it.
Import the agents you already run in Claude Code, Codex, and Cursor — identity, memory, and SKILL.md files become operator-owned data.
Durable memory, living skills, and grounded knowledge accumulate across every run — separate from any one model's context window.
Typed workflow graphs that diagnose failure, repair the plan, and judge whether a run was actually accomplished — not just "completed."
Agents author their own apps and datastores, then reach people over WhatsApp, Slack, Telegram, email, and the web.
The entire platform is one agentis.* registry agents call as code — no hidden capability, no brittle tool choreography.
Where to go next
These docs read top to bottom, in the order that matters most first. If you're ready to run something, continue to Requirements. If you want the full map before you touch a keyboard, every category is listed below in the order we recommend reading them.