Your AI agent doesn't understand your codebase.
Stop watching it read the same file four times. repowise gives Claude Code, Cursor, and any MCP agent a queryable model of your repo: architecture, ownership, and the decisions behind the code, in one call.
Your agent reads files and guesses. It does not know which files change together, which ones are dead, or why the code is shaped the way it is.
It has the source code and no memory of how the codebase got there. So it re-reads the same files, burns tokens reconstructing context you already have, and still misses the coupling and the history that would have made the answer correct. repowise gives it that memory.
Curated answers, not raw file dumps.
repowise indexes your repo once and serves it through nine task-shaped MCP tools that collapse search, read, and reason into a single call.
One call instead of four file reads
get_answer runs hybrid retrieval over the wiki, gates on confidence, and returns a cited answer. get_context returns docs, ownership, history, and decisions for any file or symbol. get_symbol returns exact source bytes without offset math. The agent asks a question and gets an answer, not a pile of files to wade through.
- get_overview for first orientation on an unfamiliar repo
- search_codebase for semantic search when you do not know where something lives
- get_risk before edits: hotspots, dependents, co-change partners
- get_why for the decision behind a piece of code
A documented dependency graph that never rots
repowise builds an LLM-generated wiki per module and file across 15 languages, rebuilt incrementally on every commit so it stays true to the code. Hybrid full-text and vector search, biased by PageRank, means your agent retrieves the right page, not a fuzzy match.
- Freshness and confidence scoring with git-informed decay
- Framework-aware edges for the major web frameworks
- Selectable wiki styles: comprehensive, reference, tutorial
- Self-hostable for private repos
Stop paying for noisy command output
Test runs, git logs, and diffs flood your agent's context with thousands of tokens it does not need. distill compresses that output errors-first before the agent reads it, and it is fully reversible if the agent wants the detail.
- 61% saved on pytest output with all 11 failures preserved
- 89% saved on git log, 86% on a 30-commit git diff
- Every error line is kept, never dropped
- A costs dashboard tallies the tokens and dollars saved
From install to first answer in four steps.
Install
pip install repowise. No cloud account required, no telemetry.
Index
repowise init builds the graph, git, health, and wiki layers from your repo.
Connect
Auto-setup for Claude Code, or register the MCP endpoint in Cursor, Cline, or Codex.
Query
Your agent calls the MCP tools and answers from a current model of your code.
Whatever you are doing, your agent has context.
Onboarding to a new repo
get_overview and the guided wiki turn a cold codebase into something your agent can reason about on day one.
Fixing a bug
get_context and get_risk surface the dependents and co-change partners so the fix does not break three other files.
Refactoring
get_health and get_dead_code rank the riskiest files and confirm what is safe to remove.
Understanding legacy code
get_why recovers the architectural decision behind code whose authors are long gone.
Reviewing a change
get_risk in PR mode returns directives: will_break, missing_cochanges, missing_tests.
Working across repos
Federated MCP queries search an entire multi-repo workspace from a single endpoint.
Your IDE's index is ephemeral, single-repo, and embeddings-only. repowise is a durable, structured layer that every agent can call over MCP, enriched with health, git history, and decisions your editor never sees, and you can self-host all of it.
Questions, answered
Which AI agents and editors does repowise work with?
repowise exposes your codebase through the Model Context Protocol (MCP), so it works with Claude Code, Cursor, Cline, and Codex, plus any other MCP-compatible client. The same index serves every agent, so you are not locked to one editor.
How much does repowise reduce my agent's token usage?
On paired benchmarks running the same model on real repositories, loading context through repowise used 96% fewer tokens (2,391 vs 64,039, about 27 times fewer), with 89% fewer file reads and 70% fewer tool calls, at answer quality on par with raw file exploration. The distill feature compresses noisy command output by a further 60 to 90 percent.
Is repowise free? Do I need to send my code anywhere?
The core engine is free and open source under AGPL-3.0. It runs 100% locally: pip install repowise, point it at your repo, and query from your agent. Bring your own API key for wiki generation, or run fully offline with a local model via Ollama. Your source code is never persisted and never leaves your machine.
How is this different from my editor's built-in codebase index?
An IDE index is usually ephemeral, single-repo, and embeddings-only. repowise is a durable, structured intelligence layer with an architecture-aware wiki, a dependency graph, git history, architectural decisions, and a defect-validated health score, all callable by any agent over MCP rather than locked to one editor.
What are the nine MCP tools?
get_overview, get_answer, get_context, get_symbol, search_codebase, get_risk, get_why, get_dead_code, and get_health. They are task-shaped: instead of forcing the agent to grep, read, and reason across many calls, each tool returns a curated, cited answer in one round trip.
How does repowise stay current as I keep committing?
It updates incrementally. A post-commit hook, file watcher, or webhook re-indexes only what changed, typically 3 to 10 wiki pages in seconds. Every MCP response carries a staleness envelope that warns the agent when the index has actually diverged from HEAD.
Which languages are supported?
Fifteen languages across the headline tiers, with full pipeline depth for Python, TypeScript, JavaScript, Java, Kotlin, Go, Rust, C++, and C#, including framework-aware route-to-handler edges for the major web frameworks.