# Prime Agent on a MicroVM: RLM Coding Without Shared Kernels

> Prime Agent wants a real machine. jurniti runs it on a Firecracker tenant microVM — BYOK, always-on, from $25/mo, 30-day money-back.

- Published: 2026-08-06 · Updated: 2026-08-06 · jurniti
- Canonical: https://www.jurniti.com/blog/primeagent-on-a-microvm

**Prime Agent** is Prime Intellect's self-improving RLM coding agent: IPython as the model tool, continual harness state, daemon-backed sessions that outlive a single TUI attach.

You still need somewhere to run it that is not "my laptop, forever open."

## What Prime Agent actually is

Upstream is open source ([PrimeIntellect-ai/prime-agent](https://github.com/PrimeIntellect-ai/prime-agent)). Install is one line:

```bash
curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh
```

Then `prime-agent` in a project directory, `/login` or a provider API key, and you work. Long-running jobs can detach and reattach (`prime-agent agents`, `prime-agent attach`).

That is a **coding runtime**, not a chat widget. It executes model-generated Python and project commands with the user's permissions. Upstream is explicit: worker/kernel isolation is lifecycle isolation, **not** a security sandbox.

So the host matters.

## Why a microVM, not a shared container slot

Prime Agent is most useful when it can:

- hold a long session (daemon + IPython state)
- refine harness state (`/refine`) without losing the machine
- keep skills and settings on disk across reboots

A multi-tenant container product shares a kernel with neighbors. jurniti's product is the opposite: **one Firecracker guest per agent**, persist on `/home/jurniti`, portal terminal for the CLI, scrub on public template capture.

You get the same isolation product as Claude Code / Hermes / OpenClaw — Prime Agent is one harness adapter among many.

## What jurniti installs for you

- Official installer path (stable channel), non-interactive at bake
- Node via nvm on the per-harness rootfs
- `prime-agent` on PATH for the login shell
- Terminal-first — no phantom web UI
- Scrub allowlist for `~/.prime/agent` (settings/skills kept; `auth.json` + sessions dropped on public templates)

You still authenticate yourself: `/login` in the TUI, or export `ANTHROPIC_API_KEY` / `OPENAI_API_KEY` / `OPENROUTER_API_KEY` / `PRIME_API_KEY` before launch.

## Who this is for

- Researchers and builders who already follow Prime Intellect / RLM work and want an always-on box
- Teams that want forkable agent setups (template → scrub → fork) without babysitting a random VPS
- Anyone who wants **Prime Agent + real isolation**, not a free shared-kernel demo

## Pricing and guarantee

From **$25/mo** (Starter). Annual available. **30-day money-back** on first purchase. Model spend is BYOK — we do not markup tokens.

No free trial. No free tier. Pay, boot, work — or get the refund.

## Related

- Landing: [Managed Prime Agent](/prime-agent)
- Upstream: [github.com/PrimeIntellect-ai/prime-agent](https://github.com/PrimeIntellect-ai/prime-agent)
- CLI coding agents compared: [cli-coding-agents-on-microvms](/blog/cli-coding-agents-on-microvms)

## Frequently asked questions

### Does jurniti offer managed Prime Agent?

Yes. jurniti provisions a Firecracker microVM with Prime Agent pre-installed via the official installer. You bring model keys or run /login in the terminal.

### Is this the same as running install.sh on a VPS?

The binary path is the same official install. The difference is tenant isolation, persist disk, snapshot/fork, and a platform portal terminal — not a shared-kernel free tier.

### Where do credentials live?

In the guest: ~/.prime/agent/auth.json or process env. jurniti never stores model API keys in the control plane.

### Cost?

From $25/mo, 30-day money-back on first purchase. Model spend is BYOK and billed by your provider.