Agent adapters

One command per machine connects a coding agent; appmarket init in a repo turns recording on there. Nothing is recorded in repos where you did not run appmarket init.

appmarket adapter install claude-code   # or: codex, opencode, cursor
appmarket doctor                        # checks each adapter against the installed agent

Agent	Prompts	Tools	Model and effort	Token usage
Claude Code	Hooks	Hooks	Session transcript	Yes, thinking tokens separately
Codex	Hooks	Hooks	Rollout file	Yes, reasoning tokens separately
OpenCode	Plugin	Plugin	Plugin	Yes, with cost
Cursor	Hooks	Hooks	Hooks	No (Cursor does not give hooks usage)
Any MCP agent	record_context		Reported by the agent

Claude Code

appmarket adapter install claude-code adds hooks to ~/.claude/settings.json (a backup is kept next to it; uninstall removes only these hooks). They record, for sessions working in a repo where you ran appmarket init and nowhere else:

• each prompt you send, and each tool call (name, plus the command or file path relative to the repo)
• at commit time, from the session transcript: model, Claude Code version, effort setting, token usage (thinking tokens separately)
• when the turn ends (the Stop hook): the agent’s final reply, as the summary of the commits it made in that turn (agents usually commit first and sum up afterwards)

When an agent runs git commit, the hook also makes the checkpoint itself, so commits are recorded even where the git hook is missing. Hook failures never interrupt Claude Code; they go to ~/.appmarket/cli.log.

Codex

appmarket adapter install codex adds three hooks to ~/.codex/hooks.json (SessionStart, UserPromptSubmit, PostToolUse; other hooks there are kept). Codex runs user hooks only after you review them once: open Codex and run /hooks. Prompts and tool calls (shell commands, and the files an apply_patch touched) come from the hooks; model, Codex version, model_reasoning_effort, token usage (reasoning tokens separately) and the final message come from the session’s rollout file at commit time.

OpenCode

appmarket adapter install opencode writes a plugin to ~/.config/opencode/plugins/appmarket.ts and adds the appmarket mcp server to ~/.config/opencode/opencode.json (other settings and MCP servers are kept; with only an opencode.jsonc, it prints the entry to add). The plugin forwards OpenCode’s events as they happen: prompts (not the task tool’s subagent prompts), tool calls with their outcome, the model and variant (effort), token usage and cost per model call, and the final message. Each event goes to a detached appmarket hook opencode, so OpenCode never waits for it. Usage of the model call that ran git commit arrives just after the commit, so it counts toward the next checkpoint. The final message becomes the summary of the commits made in that turn.

Cursor

appmarket adapter install cursor adds hooks to ~/.cursor/hooks.json (with --project: the repo’s .cursor/hooks.json; other hooks there are kept). Restart Cursor to load them. Prompts, tool calls (with their outcome), the model, the reasoning-effort setting, Cursor’s version and the final answer come from the hooks (the final answer becomes the summary of the commits made in that turn); Cursor gives hooks no token usage, so its checkpoints have none. An agent’s git commit makes the checkpoint right away, and the repo’s memory is added at session start. Hooks that fail never block Cursor (it fails open, and the hook always exits 0).

Other agents (MCP)

Agents without a hook adapter (Gemini CLI, Aider, and others for now) report their own prompt through MCP: add the server appmarket mcp (stdio) to the agent, and appmarket init adds one line to AGENTS.md (or CLAUDE.md) asking it to call record_context(prompt, summary) before each commit. These checkpoints are labelled agent-reported. When a hook adapter captured the same work, its record wins. appmarket init --agents-md adds the line even when Claude Code is present.