Quick intro, since this is my first post here: I'm Arihant, a student in India, building Aletheore alone, alongside two degrees. It's been public for about six weeks. This post is what it is, why it's different, and two real (not cherry-picked) head-to-heads against established competitors in the space.
What it does
Aletheore is a code intelligence platform: it parses a repository into a real dependency graph first, imports, symbols, API endpoints, blast radius, and grounds everything downstream in that graph instead of asking an LLM to guess from a diff or a vector-search snippet. Security scanning and PR review are two things built on top of that graph, not the whole story, there's also generated architecture documentation, dead-code detection, and semantic code search. If it can't point to the evidence, it says so instead of guessing. It ships as a CLI, an MCP server, a GitHub App that comments on PRs, and a live architecture dashboard. Free tier is real, not a crippled trial: pip install aletheore
.
Six weeks in, real numbers (not vanity metrics, just what pypistats
reports): 9,488 downloads including mirrors, 2,614 excluding them, since mid-July.
Head-to-head RepoWise generates a wiki from a codebase and answers questions against it. I ran both tools on the same 7 real open-source corpora, across 7 languages, same questions, same ground truth, best RepoWise mode shown per corpus. Full methodology and every raw result file are public: aletheore-benchmarks.
Locating code, top-1 accuracy:
| corpus | language | Aletheore | RepoWise |
|---|---|---|---|
| gin | Go | 80.0% | |
| 60.0% | |||
| serde | Rust | 53.3% | |
| 13.3% | |||
| gson | Java | 40.0% | |
| 26.7% | |||
| jekyll | Ruby | 26.7% | |
| 13.3% |
| Slim | PHP | 26.7% | 26.7% (tie) |
| guzzle | PHP | 20.0% | 20.0% (tie) |
| zod | TypeScript | 20.0% | 13.3% |
5 wins, 0 losses, 2 ties. Where we actually lose: jekyll top-5, 46.7% against RepoWise's 66.7%, stated here rather than left out. On natural-language "vocabulary" questions (phrased the way a person actually asks, not exact symbol names), RepoWise closes some of the gap since its wiki pages name the symbols directly, and on jekyll it overtakes us outright, 80.0% against our 66.7%. Across those ten additional cells: we lead in seven, tie in two, lose one.
Cost to get to a searchable index, 7 corpora total: Aletheore $0.00 (local embeddings, no API key needed), RepoWise $1.85 (LLM-generated wiki pages, $0.09-$0.47 per corpus).
Speed, measured in-process (the way an MCP server or a CLI call in a loop actually experiences it, not per-process CLI startup cost): Aletheore 40.5ms mean, RepoWise 52.5ms mean. We're faster.
Graphify is a tree-sitter-based code-knowledge-graph tool with its own query CLI. Same discipline: I ran both tools myself on frappe/erpnext (a real ~1M-LOC Python codebase), 15 independently-written questions, one shared agent loop, one anonymized judge that never knows which tool answered.
| condition | coverage | tokens/query |
|---|---|---|
| baseline (grep + read + list only) | 92.2% | 11,839 |
- Aletheore | 100.0% | 14,893 | | + Graphify | 93.3% | 17,921 |
A pre-publication review caught that one question's ground truth was wrong, and that most of the apparent coverage gap traced to a single question where the other two tools timed out without converging. Corrected for both: Aletheore and Graphify tie on coverage, and only 2 of 15 questions actually discriminate between the tools at all. The repeatable, real win is token cost: 36% fewer tokens than Graphify for the same answers.
Where we lose here: setup time. Graphify builds its whole graph on ERPNext in about a minute; when I first measured it, Aletheore's equivalent took ~23. I profiled instead of hand-waving it, found dead-code detection's own reference check was 77% of total scan time (an inefficient algorithm, not the tree-sitter parsing step I'd have guessed), and fixed it: scan time went from 236 seconds to 53, a real 4.4x, verified against the actual installed PyPI release. Setup is now ~20 minutes, with the rest in a separate indexing step that's a different, still-open problem.
Aletheore's live-docs feature covers a repo with AI-written architecture pages, but pages alone don't cover every file, obviously, no doc system covers 100% of a codebase's files by page count. Measured over Flask's last 30 real commits (100 changed files): doc pages alone had something to say about only 4 of those 30 commits' full file sets. A deterministic fallback (reads the scanner's own module record: symbols, imports, importers, no LLM call) closes that to 30/30, 100/100, at $0.00 marginal cost per file. The point isn't the doc pages, it's that a PR review tool that only works when a wiki page happens to exist isn't actually reliable, so ours doesn't depend on one.
I'll be posting more of this, real engineering, including the bugs I find in my own tool and fix in public, not a highlight reel. If you want to see it before it's public here, the dashboard has a live status page and changelog.
If you try it and something's wrong, that's exactly the kind of thing I want to hear about.