merv

Merv is a plugin for agentic coding platforms that helps agents run machine learning research as gated, reviewable experiment workflows. It works with Claude Code, Codex, GitHub Copilot CLI, Cursor, Gemini CLI, Qwen Code, Kilo Code, Hermes Agent, OpenCode, OpenHands, Replit Agent, and other MCP-capable clients. The plugin structures research at two levels. At the experiment level, each experiment starts with a generated plan that is adversarially reviewed by another agent; the plan/review loop continues until approval. After execution, the agent submits a report that is again adversarially reviewed, with the reviewer able to send work back to execution or planning. At the project level, after a set of experiments completes, five sub-agents analyze the results under different lenses. Their findings are consolidated into a report, logic graph, and change spec, which are then adversarially reviewed for accuracy. Merv consists of three main components: agent adapters that connect various agentic clients to the same workflow, a backend that owns research state (projects, claims, experiments, tasks, reviews, reflections, and workflow graphs), and a frontend for humans to inspect experiments, reviews, artifacts, logic graphs, timelines, and progress. The backend includes support systems for immutable artifacts, project feeds, agent leases, credentials, an HTTP/MCP surface, and a local runner, plus an adapter to merv-sandboxes for compute and heavy files. By default the plugin connects to a hosted brain, but the full stack can be self-hosted. In either deployment, the checkout root and caller SSH private keys remain on the user's machine; agents send explicit project IDs, typed metadata, and selected submitted bytes rather than the brain opening the checkout directly. Setup instructions are provided for Codex, GitHub Copilot CLI, Claude Code, Gemini CLI, Qwen Code, Cursor, Kilo Code, Hermes Agent, OpenCode, and remote machines. Headless runners and CI can use a static MERV_MCP_KEY.