Download Beta
Guide contents

Overview

Getting started

Concepts

Using the app

Collaboration

Reference

Supported agents

Merget commits sessions from four coding agents:

  • Claude Code
  • Codex
  • Cursor
  • GitHub Copilot

These are independent products, not part of Merget. Merget is not affiliated, associated, or partnered with — and does not sell, distribute, or get endorsed by — Anthropic (Claude Code), OpenAI (Codex), Anysphere (Cursor), or GitHub (Copilot). Merget works with the copy of each tool you have installed and licensed: it commits that tool's sessions through the tool's own hook mechanism, and can host it in a virtualized terminal inside the app. Merget never installs these tools on its own — if one is missing, it can offer to run the tool's own install command, and does so only with your approval.

For each, Merget integrates through the agent's own hook/plugin mechanism, so current versions of these agents work — there is no version matrix to check. Setup of the Merget side is automatic: when you install Merget, it detects which of the four are already installed and adds its committing plugin to each (the same thing merget plugins install does from the command line). That installs Merget's plugin into your agents — never the agents themselves. Installation is per-user, so one install covers every Merget repo you work in.

Every supported agent gets the same two things:

  1. Committing hooks — the agent reports its sessions (your prompts, its tool activity, the edits it makes) to the committing daemon as they happen.
  2. The merget skill — a document that teaches the agent to operate Merget for you. See Your agent and Merget.

What leaves your machine as part of committing is covered in Sync & privacy.

Claude Code

  • Integration: installed as a Claude Code plugin (registered in Claude Code's plugin system) plus session hooks.
  • Captured: your prompts, every tool use, file changes, and the full session lifecycle — including subagent runs.
  • Extra detail: Merget also reads the session's own log, so committed transcripts include the model used and per-message token usage.

Codex

  • Integration: installed as a Codex plugin (skill) plus global session hooks; Merget enables Codex's hooks feature for you during install.
  • Captured: your prompts, tool use, session starts (with the model in use), and turn completions.
  • Extra detail: like Claude Code, Merget reads Codex's session log for model and token-usage detail in transcripts.

Cursor

  • Integration: Cursor has no plugin activation mechanism Merget can use, so Merget installs session hooks plus a Cursor rule version of the merget skill instead.
  • Captured: your prompts, tool use, shell commands, file edits — and, uniquely among the four, the agent's responses and thinking are reported by Cursor's hooks directly.
  • Not captured: Cursor does not report which model a session used, and Merget does not read Cursor's session log today, so token detail is limited compared to Claude Code and Codex.
  • Compatibility note: one of Cursor's local settings conflicts with Merget's hooks. The Merget app checks for this and, if needed, shows a Cursor Compatibility dialog offering a one-click fix — it updates the setting and restarts Cursor (save your work in Cursor first). Accept it; Cursor sessions aren't committed reliably until the fix is applied.

GitHub Copilot

  • Integration: installed as a Copilot plugin plus session hooks (this is the Copilot CLI coding agent).
  • Captured: your prompts, tool use and failures, errors, subagent runs, and the full session lifecycle. The model is captured when Copilot reports it.
  • Not captured: Merget does not read Copilot's session log today, so token detail is limited compared to Claude Code and Codex.

Other tools and editors

Tools outside these four still get file-change committing: the committing daemon watches the repo's files, so edits made with any editor or unsupported AI tool are captured as steps attributed to you. What you don't get is intent — no prompts, no transcripts, no agent attribution. See How Merget works.

Under the hood

Each agent exposes a hook API that fires events at session milestones — session start, prompt submitted, tool started/finished, turn end. Merget registers a small hook command for the events each agent supports, and those events are what the committing daemon assembles into steps with transcripts. The per-agent differences above are exactly the differences in what each agent's hook API reports: Cursor's hooks include agent responses and thoughts but no model name; Claude Code and Codex additionally write a local session log that Merget tails for token usage. The skill is the same document everywhere — Cursor just consumes it as a rule instead of a plugin skill.

Next: Excluding files · Related: Your agent and Merget · Terminals