Skip to main content

​
What the plugin installs

  • Eight skills — the operating loop and the review protocol (research-workflow, experiment-design-review, experiment-attempt-review), reflection (project-reflection, project-reflection-review, consolidation-review), sandbox hygiene (sandbox-operation), and posting (feed-posting).
  • Four reviewer agents where the client supports them (namespaced merv: on clients that namespace agents).
  • The MCP connection to the brain — 41 tools by default, schemas served live. MCP tool contract
How each arrives per client: Clients.

​
What to add

  • A project instruction (CLAUDE.md, AGENTS.md, .cursor/rules, …) that names your project and says: “Use Merv for research. Start every research task with workflow.status_and_next. Never complete an experiment outside MCP.” Keep it short — the skills carry the detail.
  • The Nisa skill, if you want literature search in the same agent: Using Nisa from a coding agent.
  • For OpenHands or Replit Agent, paste the skill text you need into the agent’s instructions; nothing is auto-installed there.

​
Prompting that works

  • State the claim to test and the constraints (data, compute, time), not the procedure. The workflow supplies the procedure.
  • Ask for the smallest experiment that gives a decision-relevant signal, and for the decision rule to be written before the run.
  • Say “request a review” rather than “check the plan” — the former triggers the real handoff.
  • Ask the agent to post to the feed as it goes; it makes the session readable later.
  • When it stalls, ask it to call workflow.status_and_next and read next_action and missing_evidence back to you.

​
Reviewer isolation

A different session from the producer, always — natively on clients with subagents, a second session elsewhere. Reviewer handoff

​
Auto-run

To have experiments run without you, pair a machine: Auto-run. Your interactive agent and the runner can share a project.