Skip to main content
Merv is a plugin for coding agents — Claude Code, Codex, Cursor, and eight more. It gives a research agent structure: plan an experiment, get the plan reviewed, run it, get the results reviewed, and periodically step back and reflect on the whole project. You install it in the agent you already use. The agent talks to a hosted brain over MCP; a web UI shows everything from research strategy to a single run’s output.

​
Every experiment passes two reviews

Experiment workflow: Plan, Design review, Execute, Results review, Complete. Rejected reviews send work back to Execute or Plan.
A second agent reviews the plan adversarially before anything runs; a third reviews the report against the plan afterwards and can send it back to execution or to planning. Nothing skips a gate. The gated workflow

​
Every few experiments, the project reflects

Project workflow: completed experiments fan out to three core lenses plus two written for the wave, then Synthesis, Reflection review, Publish.
Five sub-agents read one snapshot — three fixed lenses (what to amplify, what to avoid, what hasn’t been tried) plus two written for that wave — and their findings become a reviewed report, project graph, and change spec for the next phase. Reflection waves

​
Three pieces

System architecture: agent platforms connect to the brain over authenticated HTTP MCP; the brain owns records, gates and sandboxes; the frontend supervises the brain.
Hosted by default; self-host if you want. Your checkout and your SSH keys stay on your machine. Agents send explicit project ids, typed metadata, and the bytes they choose to submit; the brain never opens your repository.

Choose your setup

Hosted or self-hosted; interactive or auto-run.

Quickstart

Install, sign in, first call — five minutes.