Codebase intelligence · Open source
Repowise maps structure, git history, tests and decisions into one index that stays current on every commit. Agents stop guessing, pull requests show what they can break, and the riskiest code comes with a named fix.
repowise-dev/repowise · 5,607 pages, rebuilt on every commit
Measured, not claimed
Used by developers at
Engineers who use Repowise for their projects work at these companies.
Why AI velocity stalls
AI made producing a change cheap. Understanding what it touches is still expensive, and someone pays that cost on every task.
01
Your agent reads three files and confidently changes a function that forty callers depend on. Nothing told it.
02
The diff took seconds to write. The review takes the afternoon, rebuilding enough context to judge it.
03
The retry logic is shaped by an outage two years ago. The engineer who knew left. Last week an agent simplified it.
One index. Three jobs.
Repowise connects structure, history, tests and decisions once. The same evidence then answers three different questions.
0.876 · file coverage on 42 sealed tasks
Every file, module and decision gets a page that rebuilds itself on each commit. Ask in plain English and get the lines the answer came from.
handle_event
fresh · 2 min agoVerifies the signature, then applies the event exactly once. Retries are safe because the event id is recorded before the ledger write.
def handle_event(event): verify(event.signature) if ledger.seen(event.id): return ledger.apply(event)Why · decision 14
Idempotent by event id. Added after the double-charge incident; do not remove the check.
cited · hooks.py:41-58 · ledger.py:12
0 · LLM calls per pull request
Before a change merges, see every caller it reaches, the files that usually move with it, and the tests that prove it still works.
billing/ledger.py · changed together 9 of 11 timestests/test_hooks.py, tests/test_replay.pyrepowise bot · on the PR
Reaches 6 files. ledger.py usually changes with this one and is missing.
0.74 · bug-prediction AUC across 21 repos
Every file is scored on defect risk, maintainability and performance. Repowise names the fix, not just the file.
Defect risk
7.4/ 10
Good · up 0.6 this month
payments/processor.py3.9payments/routes.py6.2payments/models.py8.1orders loop reaches db.get_item in 2 hopslegacy_refund() · high confidence, safe to deleteFix first · extract method
Split settle() into 3 helpers. Behaviour checked by dataflow; run tests/test_settle.py.
handle_event
fresh · 2 min agoVerifies the signature, then applies the event exactly once. Retries are safe because the event id is recorded before the ledger write.
def handle_event(event): verify(event.signature) if ledger.seen(event.id): return ledger.apply(event)Why · decision 14
Idempotent by event id. Added after the double-charge incident; do not remove the check.
cited · hooks.py:41-58 · ledger.py:12
Also in the indexOwnership and bus factor, secrets across the whole git history, docs that drifted from the code, and contracts between repositories.
Coming to hosted and enterprise
Trained models that learned from years of real commits and fixes. Paste a bug report and Foresight ranks the files the fix will touch, each with a reason. No LLM, milliseconds on a CPU.
Issue #4127 · open
Checkout sometimes charges twice after a timeout
On a slow connection the payment page times out, the customer tries again, and the card is charged twice.
payments/retry.pyfix lands here
41def charge_with_retry(order):42 try:43 return charge(order, key=new_key())44 except Timeout:45 return charge(order, key=new_key())The retry asks for a new idempotency key, so the second attempt becomes a second charge.
Keyword searchForesight · files the fix will touch
search rank
payments/retry.pyfixed for a similar timeout bug 3 weeks ago.71#9billing/ledger.pychanges with retry.py 9 of 11 times.14#7payments/gateway.pyraises the Timeout that retry.py catches.06#6web/checkout_page.tsxmatches “checkout” 14 timesUI only, no payment calls.02#1tests/test_checkout_ui.pymatches “checkout” 9 timestests the page, not the charge.01#2README.mdmatches “charges” 3 times.01#3web/checkout.cssmatches “checkout” 6 times.01#4docs/checkout.mdmatches “timeout” 2 times.01#5Same evidence, every surface
One index behind your agent, your pull requests, your editor and your team's dashboard. Nothing new to switch to.
Your agent calls it over MCP
Ten MCP tools, session hooks and a generated CLAUDE.md, so the agent starts with the map.
One comment per PR with callers, missing tests and risk, and a merge check beside it.
File health in the status bar, blast radius from source control, decisions on hover.
Architecture, owners, hotspots, decisions and health in one place, shareable by link.
Two ways in
Give your agent the whole repo, and keep the map current without thinking about it.
One shared map of every repository, deployed where your code already lives.
What we publish
Every number on this site ships with its sample size, its conditions and the raw data. The benchmarks page includes the measurements where another tool wins.
Read the full recordRepowise analyzes source structure, dependencies, call relationships, git history, ownership, code-health signals, tests, and engineering decisions. It turns that evidence into a reusable codebase model and generated documentation.
Yes. You can open and browse supported public repositories without an account or installation.
The Free hosted experience includes 10 grounded answers each month. Those answers use indexed repository evidence instead of treating the codebase as an unstructured prompt.
No. Repowise supplies codebase context to developers and existing tools through the hosted application, CLI, and Model Context Protocol. Your agent or editor remains where the work happens.
Yes. The core engine is open source under AGPL-3.0 and can run locally with your model provider or a local model. Repowise is also available as a hosted product, with an enterprise path for organizational deployments.
Start with a real repository
Paste any public repository, or install the PR bot on your own.
Evidence you can inspect
Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.
The library for web and native user interfaces.
The agent engineering platform.
