Choose your setup

Where the brain runs

Brain	Hosted (default)	Self-hosted
What you install	The Merv plugin in your coding agent	The plugin, plus the brain stack (brain, Postgres, S3-compatible store)
Sign-in	Browser OAuth from the agent; no keys to copy	Same clients, same OAuth — you point the MCP `url` at your brain
Sandboxes	Provisioned by the hosted brain with your provider credentials	Provisioned by your brain
Best for	Getting started, small teams	Data-residency requirements, air-gapped labs, hacking on Merv itself

Same open-source code either way. Self-hosting starts at Self-host → Overview.

How agents run

Agents	Interactive (default)	Auto-run
What runs the agent	You, in Claude Code, Codex, Cursor, … with the Merv plugin	A paired machine running the Merv runner, dispatching agents from a queue
Install	One plugin command per client — Quickstart	One installer command, then a pairing code in the UI — Auto-run
Sign-in	Browser OAuth	Pairing code — the machine mints its own key locally, only its digest is sent to the brain
Best for	Working alongside the agent, reviewing as you go	Letting experiments run overnight or across a fleet

You can use both: work interactively on your laptop and pair a workstation or a remote box for auto-run. See Auto-run.

Where experiments execute

Wherever the agent is, an experiment can run:

Locally, in the agent’s checkout, for cheap runs.

In a cloud sandbox the brain provisions on your provider account (GPU or CPU, with a lease). The agent runs commands over SSH and pulls outputs back. See Sandboxes.

If you’re new: hosted brain, interactive agent, local runs. Add sandboxes when a run needs a GPU, and auto-run when you want experiments to keep going while you sleep.

Start

Concepts

Guides

Clients

Reference

Self-host

Where the brain runs

How agents run

Where experiments execute

​Where the brain runs

​How agents run

​Where experiments execute

Where the brain runs

How agents run

Where experiments execute