Set up with your agent

You do not have to set Repowise up by hand. Paste this into your coding agent:

That page is a runbook written for agents. It works in Claude Code, Codex, Cursor, VS Code, OpenCode, Hermes, and any other agent that can fetch a URL and run shell commands.

What your agent will do#

1. Check for Python 3.11+ and a git repository, and install the CLI with uv, pipx or pip if repowise --version fails. If the repo already has a .repowise/ index, it runs repowise status and does not re-index.
2. Index the repo with repowise init --yes --no-prose. This builds every layer and renders the documentation from the code's structure, with no model and no spend, even if an API key is set in your environment.
3. Connect itself: the right init flags or repowise agents add command for the host it is running in.
4. Ask you once whether to add an LLM and semantic search (below).
5. Confirm the MCP tools are live by calling get_overview, and run repowise doctor if they are not.
6. Report back with the architecture, the top hotspots, and the health findings worth acting on.

What it asks#

Everything works keyless. The agent then recommends two additions and asks once:

• An LLM. get_answer writes cited answers instead of returning ranked evidence, the docs explain why the code is shaped the way it is, and decisions are mined from the history. In Claude Code or Codex the agent offers your existing subscription (claude_cli, codex_cli, copilot_cli), so there is no API bill. With an API key it shows the repowise generate --dry-run estimate and spends nothing until you approve. It never asks you to paste a key into the chat.
• Semantic search. Search and get_answer match by meaning, not only by words. It needs a Gemini, OpenAI or OpenRouter key, or Ollama locally.

Say no to both and nothing leaves your machine. The graph, git history, code health, change risk and dead-code layers are the same either way.

If a key is already in your environment, the agent names it and still asks before spending. Subscription providers (claude_cli, codex_cli, copilot_cli, opencode) and local Ollama need no key. An agent CLI has no embedder, so semantic search needs an embedder key of its own. The full provider list is in the runbook.

Already installed?#

If Repowise is on your machine and you only want the repo indexed:

Run `repowise init --yes --no-prose` in this repo. It needs no API key and
spends nothing. Then call get_overview and summarize what this codebase does.

Install the plugin#

The fullest setup is the plugin. One install wires the MCP server, six auto-activating skills, 18 slash commands, and the context hooks together:

/plugin marketplace add repowise-dev/repowise
/plugin install repowise@repowise

There is a Codex plugin too. Both are covered in Agent plugins, including what each skill activates on and what the hooks do.

If you only want the MCP server and none of the rest:

claude mcp add repowise -- repowise mcp

Or commit a project .mcp.json so your whole team picks it up:

{ "mcpServers": { "repowise": { "command": "repowise", "args": ["mcp"] } } }

Other agents#

Eight agent targets are wired end to end: repowise agents add --target=<id> writes the MCP config and, where the host reads one, a managed instructions file. The tier is derived from what a descriptor actually implements, not declared:

Full means MCP tools, a managed instructions file, hook-level interception of tool calls, and transcript mining after the fact. Good means MCP tools and a managed instructions file where the host reads one, without transcript mining: the agent can ask repowise questions, but repowise sees little of what happens in between. Cursor also gets a preToolUse distill hook and an afterFileEdit hook. Attribution means repowise can record the agent's edits as they happen, an opt-in set by attribution.capture (see repowise attribution); every agent's git-ai notes, Agent Trace files and commit trailers are read either way. Target ids for --target=: claude-code, codex, vscode, cursor, opencode, hermes, copilot, kiro.

codex mcp add repowise -- repowise mcp

repowise init --codex

repowise agents add --target=cursor

{
  "mcpServers": {
    "repowise": {
      "type": "stdio",
      "command": "repowise",
      "args": ["mcp", "/path/to/your/project"]
    }
  }
}

repowise agents add --target=opencode

repowise agents add --target=hermes

repowise agents print-config claude-code

{
  "mcpServers": {
    "repowise": {
      "command": "repowise",
      "args": ["mcp", "/path/to/your/project"]
    }
  }
}

What agents should know about the CLI#

The behaviors that most often trip up an unattended run, all of them deliberate:

Repowise also writes a managed block into .claude/CLAUDE.md and AGENTS.md during init, so agents that read those files get the tool guidance without any MCP call at all. The block is lean by default (date and commit, the files that need care, one code-health line, the top three decisions, three working bullets and the tool table), and editor_files.sections adds sections back.

Machine-readable docs#

This documentation site serves itself in agent-friendly form:

Start with llms.txt to find the right page, then fetch that page's .md, or take llms-full.txt when you want everything at once. Every page also has a Copy page control at the top, which copies its Markdown or opens it in Claude or ChatGPT.

What to do next#