Stop Making Your AI Agent Re-Learn Your Codebase Every Time

Every task it re-discovers your codebase from scratch.
grep, read, re-read, forget, repeat.
It burns most of its budget BEFORE it writes a single line.
Repowise fixes that. 👇
⚡️ Index once. Never rediscover again.
One index, kept current on every commit. Your agent reads the answer instead of the codebase.
📉 The numbers:
→ 64,039 tokens → 2,391 tokens to load one commit's context [get_context]
→ -96% tokens, -89% file reads, -70% tool calls to load context
🗺️ What one index builds - 5 layers:
→ ◈ Graph: Dependency graph across 17 languages, symbol-level, with PageRank & call resolution
→ ◈ Git: Hotspots, ownership, hidden coupling (co-change), bus factor, bug history
→ ◈ Docs: Auto-generated wiki + auto-generates your CLAUDE.md & AGENTS.md from the REAL index
→ ◈ Decisions: Architectural decisions mined from 8 sources
→ ★ Code Health: 25 deterministic markers - 1 to 10 per file, zero LLM, under 30s on a 3,000-file repo
🎯 It names the ACTUAL fix:
Not "this file is risky". It says Extract Class / Move Method / Break Cycle / Split File, with exact methods + blast radius attached
Plus 10 task-shaped MCP tools for Claude Code, Cursor, Codex.
✅ Free, open source (AGPL-3.0), runs 100% on your machine. No API key for first index.
pip install repowise

Here's what I can verify from Repowise's own site, GitHub, and PyPI listings — with the caveat that these are all self-reported by the project, not independently audited.

Checks out (matches the vendor's own materials):

• It's real: pip install repowise, open source, AGPL-3.0 license, self-hosted, no API key needed for a basic index. get_overview and get_context work in index-only mode with no key, synthesized from the graph, git and health layers.
• Ten MCP tools, not the marketing post's implied set — the actual toolkit is get_overview, get_answer, get_context, get_symbol, search_codebase, get_risk, get_change_risk, get_why, get_dead_code, get_health.
• "Index once, kept current on every commit" — plausible: It updates incrementally. A post-commit hook, file watcher, or webhook re-indexes only what changed, typically a handful of wiki pages in seconds, and every MCP response carries a staleness envelope that warns when the index has diverged from HEAD.
• Docs/wiki auto-generation and CLAUDE.md-style output is a documented feature, not invented.
• AGPL-3.0, local-first, no forced API key: The core engine is open source under AGPL-3.0 and runs 100% locally: pip install repowise, bring your own API key, or run fully offline with a local model.

Flags / discrepancies worth knowing:

• Language count is inflated in the post. The post says "17 languages." Repowise's own site says Fifteen languages across the headline tiers, with full pipeline depth for Python, TypeScript, JavaScript, Java, Kotlin, Go, Rust, C++, and C# — 15, not 17, and only 9 of those get "full pipeline depth."
• "5 layers" vs official docs are consistent (graph, git, docs/wiki, decisions, code health) — that part checks out against Repowise indexes your codebase into five intelligence layers (graph, git, wiki, decisions, code health) and exposes them to AI agents through ten task-shaped MCP tools. Note the GitHub README describes only four layers in one spot (dependency graph, git, docs, decisions) — health scoring seems to be a newer fifth addition, so the docs site is more current.
• The token-reduction numbers (64,039 → 2,391 tokens, -96%) aren't in the sources I could verify, and the closest official benchmark I found is a much more modest, hedged claim: On 48 paired SWE-QA tasks from pallets/flask, repowise-augmented Claude Code matches baseline answer quality while being dramatically leaner: 32 / 48 (67%) tasks are cheaper with repowise, at parity quality. That's 67% of tasks being cheaper — not a universal 96% reduction. The specific numbers in the post may come from one cherry-picked example rather than an average.
• Init isn't free-as-in-instant. Full indexing with wiki generation costs money and takes real time: Full, uses your chosen LLM to generate human-readable wiki pages for every file and module. Advanced, fine-tune every option (concurrency, exclusions, commit limits, etc.) A typical first run on a medium codebase (~500 files) takes 5-15 minutes and costs $1-5 depending on the provider. The post's "no API key for first index" is true only for the index-only mode, not the full doc-generation mode it's advertising.
• "Code Health: 25 deterministic markers... zero LLM" is consistent with their positioning as a CodeScene alternative with a defect-validated health score (ROC AUC 0.74 across 21 repos) — that's a real, specific stat, though I can't verify the methodology behind the 0.74 AUC figure independently.

Bottom line: The product is real and the core mechanics (MCP tools, incremental re-indexing, five layers, open source) are accurately described. The two numbers most likely to be misleading are the language count (17 vs. the vendor's own "15") and the headline token-reduction stat, which doesn't match any published benchmark I could find — the actual published number is "67% of tasks cheaper," not a flat 96% reduction.