With AI coding agents

With AI coding agents

pipx install research-diff
pipx inject research-diff mcp
research-diff setup claude          # --scope project or local for one repository; --undo removes it all
research-diff watch                 # http://127.0.0.1:8766/

setup claude adds three things and prints each:

• a Stop hook, so that after every turn one line says what changed in the research;
• a skill, which tells the agent to write markers, log results and show you the diff;
• the MCP server.

The loop

1. You and the agent work on an idea: notes, a config, a run, a paragraph of the paper.
2. After the run, the agent records what came out with log_result: the metrics and what they support or weaken. It lands in research/log.md as an R marker, so it is on the map, in git and in the next diff.
3. The hook's line and the live page show what the turn did, e.g. H1: open → refuted; 1 new warning (C1 relies on H1, which is refuted).
4. Ask the agent to explain. It calls research_diff and answers from the exact plan.
5. Mark as seen (or research-diff checkpoint, or ask the agent): the next diff starts there.

The MCP tools

Tool
research_diff(base="checkpoint", head=".")	the headline, the plan, new warnings and the picture's path
research_map(at=".")	every node with its status and metrics, the links, the warnings and the picture
log_result(experiment, metrics, title, supports, weakens, note)	records a run as the next R marker in research/log.md
checkpoint(ref="")	marks what you have understood; the skill tells the agent to call it only when you say so

The agent never marks a hypothesis supported or refuted on its own: the skill tells it to ask you.