Start Here · 01.02
Agentis Requirements: What You Need Before You Install
Agentis needs Node 20.10+ and pnpm to start — embedded SQLite plus a pinned, verified local embedding runtime.
The short answer
The installed Agentis CLI needs Node ≥ 20.10 and a modern browser. If you are running Agentis from this source repository, use its pinned pnpm workspace too. It ships with embedded SQLite and a content-addressed local asset store. The Brain's pinned q8 ONNX artifacts (about 129 MB) are downloaded and SHA-256 verified on first warmup, then semantic recall runs offline with zero model-provider API keys. Attach a model runtime when you are ready to let an agent execute turns.
What's bundled vs. what you configure
| Component | Default | Override |
|---|---|---|
| Runtime state | ./.agentis (SQLite, secrets, agent homes, backups) | AGENTIS_DATA_DIR |
| HTTP server | 127.0.0.1:3737 | AGENTIS_HTTP_HOST / AGENTIS_HTTP_PORT |
| Embeddings | pinned local ONNX (multilingual-e5-small q8), verified once and reusable offline | workspace embedding provider config |
| Assets | content-addressed blob store, deduped by SHA-256 | AGENTIS_ASSETS_DIR |
Every row above works without an account or credit card. A fresh local Brain needs network access once to fetch its pinned model, or a cache prepared elsewhere with agentis warmup.
What you'll want to add — and when
Agentis boots complete, but an empty workspace can't think yet. The one thing you attach next is a model runtime: a CLI harness (Claude Code, Codex, Cursor), a local model, or any OpenAI-compatible HTTP endpoint. Everything else below is genuinely optional, added only when you need it:
| You want to… | Then attach |
|---|---|
| Give an agent a model to think with | One CLI harness or HTTP endpoint — see Bootstrap Your First Agent |
| Reach WhatsApp, Slack, Telegram, or email | A channel connection — see Reach |
| Generate images, audio, or video | A media provider key — see The Agent-Native SDK |
| Export traces to an observability backend | AGENTIS_OTEL_ENDPOINT — falls back to a no-op tracer if unset |
Storage and platform notes
SQLite is single-writer by design — this is an intentional choice for a local-first, single-tenant operator deployment, not a scaling shortcut you'll hit later. Plan disk space around three things that grow over time: the hot SQLite database, the cold archive directory under AGENTIS_ARCHIVE_DIR, and the content-addressed asset store under AGENTIS_ASSETS_DIR, which holds every image, document, and generated artifact your agents produce.
Agentis keeps active runs fully hot, then moves completed run state, ledger events, and observability rows into compressed cold archives after the configured retention window. Reads hydrate archived data back into run history and trace views, so operators keep the evidence without letting a young workspace turn into an oversized hot database.
Agentis runs on Windows, macOS, and Linux. The examples throughout these docs use POSIX shell syntax; swap in the PowerShell equivalent if that's your environment.
Continue
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.
Install the Agentis CLI globally, run agentis up, and open the local dashboard on 127.0.0.1:3737.