Poolside trained its Laguna models inside an agent harness. That harness is pool — a terminal TUI, an ACP server for editors, an ACP client for other agents, and a headless pool exec for CI. Hosting "just any coding agent" with Laguna is fine; hosting pool is what the models were optimized for. Here's the clean setup: run pool on its own always-on Firecracker microVM. Give it root there, wire MCP and AGENTS.md, and a bad tool call stays contained to a box you can rebuild — booting in about 124ms.
What pool is
pool is Poolside's coding agent. Install once:
curl -fsSL https://downloads.poolside.ai/pool/install.sh | sh
Then run pool in any project. Interactive mode has slash commands, @ file fuzzy search, shell mode with !, and mode cycling (Always ask / Accept edits / Allow all / Plan). Spec support is first-class: AGENTS.md, Skills, MCP, and ACP.
It also speaks agent protocols outside the TUI:
pool acp— ACP server for Zed, JetBrains, Xcode, and any ACP clientpool --agent-server …— drive Claude Agent, Codex, or remote ACP over HTTPpool exec— one-shot non-interactive runs for scripts and CI
Auth is flexible: pool login (including OpenRouter) or POOLSIDE_API_KEY for automation. Config and credentials live under ~/.config/poolside/ (settings.yaml + credentials.json).
Why a real microVM for pool
pool acts. It runs tools, edits the workspace, and can connect to remote MCP servers with headers and env that hold secrets. On a laptop, that blast radius is your whole home directory. A shared container still shares the host kernel.
jurniti gives pool its own Firecracker microVM — its own kernel, a KVM hardware boundary, one tenant per box. Give pool room to work inside its own VM and a runaway command, a misbehaving MCP server, or a prompt-injection stays on that box. Reprovision and you're clean in seconds.
You keep your model relationship. jurniti's BYOK is architectural: pool login or export POOLSIDE_API_KEY=…, and the key lives only in your VM. Traffic leaves straight for Poolside inference or OpenRouter; jurniti is never in the path, and no token is marked up.
Most pool work runs comfortably on Starter (1 vCPU / 2 GiB). And it's reversible: a 30-day money-back guarantee, workspace kept 7 days after cancellation.
Run pool in 3 steps
By hand this is a VPS, the install script, EULA accept, PATH, login, MCP config, and a unit file you maintain forever. jurniti runs the box; you keep the agent.
1. Pick a plan and pay
Starter is enough for pool. Check out, and the provision chain runs automatically the moment payment confirms.
2. Your microVM boots with pool ready
A fresh Ubuntu microVM boots with the pool binary pre-installed and on your PATH. You get an in-browser terminal — nothing to SSH into. Authenticate and go:
pool login # Poolside account or OpenRouter
# or: export POOLSIDE_API_KEY=...
pool # interactive TUI in the current project
3. Put it to work
Hand pool a task, add MCP servers with pool mcp add, drop an AGENTS.md, and let it run. Settings and project wiring live on the persistent home, so they survive every restart.
What people run pool for
- Laguna-native agentic coding. Use the harness the models were RL-trained in, not a generic shell wrapper.
- ACP in the cloud. Point an editor ACP client at a remote microVM, or run
pool acpas the server. - Headless CI jobs.
pool exec -p "…" -o jsonon an always-on box without parking a laptop. - MCP-heavy workflows.
settings.yamlMCP servers and permissions, scrubbed cleanly when you publish a template.
Snapshot a configured agent into a template and fork it for the next job — your settings and AGENTS.md, never your credentials.
jurniti vs a laptop vs a DIY VPS
| jurniti | Your laptop | DIY VPS | |
|---|---|---|---|
| Always on | Yes | No — it sleeps | Yes |
| Isolation for an agent that runs tools | Firecracker microVM (KVM) | Your whole machine | Shared kernel |
| Provider keys | Yours, never leave the VM | Yours | Yours |
| MCP / tool blast radius | One disposable microVM | Your whole machine | Your whole server |
| Setup | Automatic · ~3 min to live | Local install | Manual |
| Refund | 30-day money-back | n/a | None |
Pricing
Starter is $25/mo ($250/year, two months free) and runs pool for most work. Pro is $49/mo and Max $99/mo for heavier repos or more concurrent tool use. Every plan includes the 30-day money-back guarantee, and your model spend is separate and never marked up.
Running a different agent? See the guides for Goose, Claude Code, and OpenCode, or compare plans on the pricing page.
