For AI agents

Before your agent edits, it knows what the change breaks.

Repowise indexes your repo once: structure, history, tests and decisions. Your agent reads that index through ten Model Context Protocol (MCP) tools and hooks, so it sees the callers, owners and tests of a file before it writes. Open source, local, no model key.

pip install repowise

Works with

Before the edit: the index answers with the callers, the tests, the owner and the decision behind the file.
After review: a finding on the diff arrives as a prompt your agent applies.

What your agent gets

Everything it was guessing at, one call away.

  • Ten MCP tools

    get_answer, get_context, get_risk, get_why and six more. Each does a whole task and names the commit it describes.

  • Hooks that arrive first

    Context lands at session start, on broad searches and on stale reads. A bug-magnet file gets a warning mid-edit.

  • Distill

    Test, build, git and search output, errors first. 89% fewer tokens on git log -50.

  • A generated CLAUDE.md

    repowise init writes CLAUDE.md or AGENTS.md with the architecture, the files that need care and the standing decisions.

  • The tests to run

    get_risk names the tests a change reaches, straight from the graph. 97.5% precision on the repowise repository.

  • Local and keyless

    Graph, history, health and dead code are computed on your machine with no model call, and update on every commit.

31.6%

fewer output tokens in a real agent loop

n = 43 · p = <0.0001 · Codex (gpt-5.6-sol) · django/django at a single pinned commit · 2026-08

3.8

Tool calls to reach an answer, against a bare agent's 7.2

n = 43 · Codex (gpt-5.6-sol) · django/django at a single pinned commit · 2026-08

89%

Tokens saved distilling git log -50

n = 1 · one command on one repository · 2026-08

MCP tools

One call answers what ten file reads would guess at.

Each tool returns the evidence for one job: who depends on a file, how often it broke, who owns it, which decision governs it. Every response names its commit, so the agent knows how fresh it is.

  • get_answer, get_context, get_symbol, search_codebase, get_why, get_risk, get_change_risk, get_health, get_dead_code, get_overview
get_risk, MCP response (excerpt)
{
  "targets": {
    "packages/core/src/repowise/core/pipeline/persist.py": {
      "dependents_count": 43,
      "hotspot_score": 0.9989,
      "co_change_partners_total": 25,
      "primary_owner": "Raghav Chamadiya",
      "bus_factor": 1,
      "defect_profile": {
        "fix_count": 39,
        "bug_magnet": true,
        "window": "6 months"
      },
      "line_coverage_pct": 85.47
    }
  },
  "_meta": {
    "indexed_commit": "e07c79f151ee",
    "index_behind": false
  }
}
One call on a file in the repowise repository. Fields trimmed, values as returned.

Hooks

Context arrives before the agent asks for it.

In Claude Code and Codex, a session starts with the index freshness and the decisions on the branch. A broad search gets the related files and their owners. A file that changed since the agent read it is flagged as stale.

  • Local reads only: no model calls, no network
  • Edit-time notice when a decision governs the file
SessionStart hook, Claude Code
[repowise 0.56.0] Codebase index is current (HEAD e07c79f1). The MCP tools (get_answer, get_context, search_codebase, get_risk) serve content verified against this exact tree; prefer them over raw file reads for locating and understanding code before you edit.
[repowise] Proposed but NOT accepted - mined from prior sessions and awaiting review. Weigh these; do not treat them as rules:
- Make the composed refactoring opportunity the unit on every surface
What the agent reads before its first message. The decision line is one of the repository's own.

Where it runs

One index for every agent you use.

repowise init wires up the agents it finds and writes their instructions file. On hosted Repowise the same tools answer over a URL, per repository or across a workspace.

  • Claude Code and Codex at the full tier: tools, hooks, skills and an instructions file
  • Cursor, VS Code, OpenCode, Copilot CLI, Kiro and Hermes set up by repowise init
  • Windsurf, Zed, Cline, Gemini CLI and any MCP host with one pasted entry
$ repowise agents
Agent        Tier   Status
claude-code  full   configured
codex        full   configured
vscode       good   configured
cursor       good   installed, not configured
opencode     good   installed, not configured
hermes       good   installed, not configured

Add or remove with repowise agents add --target=<id>.
The agents found on one machine, and which ones are wired to the index.

FREQUENTLY ASKED

Questions, answered

Which agents does repowise work with?

Claude Code, Codex, Cursor, VS Code, OpenCode, GitHub Copilot CLI, Kiro and Hermes are set up by repowise init or repowise agents add. Windsurf, Zed, Cline, Gemini CLI and any other client that speaks MCP take one pasted server entry, locally or over the hosted endpoint.

Does my code leave my machine?

Not for the analysis. The graph, git history, code health and dead code are computed locally with no model call. Generated prose is optional and goes to the model provider you configure, under your own key.

What does it cost?

The CLI and MCP server are free and open source under AGPL-3.0. Hosted Repowise is free on public repositories; Pro adds private ones.

What are the ten tools?

get_answer, get_context, get_symbol, search_codebase, get_why, get_risk, get_change_risk, get_health, get_dead_code, get_overview. Start with get_overview on a new repository, then get_answer or get_context for a question and get_risk before an edit.

How does the index stay current?

repowise init installs a post-commit hook that updates only what changed. Responses carry the indexed commit, and the session hook says when the index is behind.

Get started

Install it and connect your agent.

One install, one init, and your agent has the repository.