local · offline · cited
Reify reads a brownfield codebase — source, SQL, business documents and fifteen years of git history — and compiles it into a queryable model. Your agent asks for a task and gets back the smallest set of facts that answers it, every one carrying a file and a line number.
❯ curl -fsSL https://raw.githubusercontent.com/lambiengcode/reify/main/install.sh | sh ❯ reify init && reify index ❯ reify context "add a discount tier for strategic customers" --budget 1500
SWE-bench Verified is 500 real GitHub issues. For each one the index is built at that issue's own base commit, so the fix cannot leak into what Reify is allowed to know. Both arms get the same token budget.
Retrieval is not the same as resolution, and the README says so with equal prominence: end to end, Reify and a BM25 baseline currently resolve the same number of issues. The full write-up shows both, including the arms that failed.
Six commands, each answering a question an agent otherwise guesses at.
The flagship. Compiles the minimum useful context for a task, under a token budget you set.
For one line: its callers, the tables it writes, the files it co-changes with, and the commits that explain it.
What breaks if this changes, including hops no single grep would follow.
One business concept across every language, table and file it appears in.
Documentation that disagrees with the implementation, with both sides cited.
The Model Context Protocol over stdio, three tools, for agents that speak it.
Reify never states a guess as a fact. Each claim is tagged with how it is known, so an agent — or a reviewer — can tell a parsed fact from an inference without reading the docs.
[confirmed] erpnext/accounts/doctype/journal_entry/journal_entry.py:492 calls check_credit_limit [observed] customer (is written by after_insert) [inferred] journal entry check credit limit: if customers from …customer import check credit limit [unknown] no document section describes this symbol
Parsed directly from source. A fact.
Seen in behaviour or history, repeatedly.
A lead to verify against its citation, never a fact.
Two sources disagree, and both are shown.
Nothing is known — said plainly, not guessed.
Not "by default" — at all. There is no HTTP client in the dependency tree, and the test suite fails the build if one appears. For a company that will not let proprietary code near a cloud service, that is the difference between a tool they can evaluate and one they cannot.
Networking crates in Cargo.lock | asserted zero, in CI |
|---|---|
| Sockets in the source | asserted zero, in CI |
| Subprocesses | git, reviewed document converters, and — for reify upgrade only — curl and tar; each named in a test |
| Telemetry | none, and no flag that turns any on |
No protocol, no server, no per-turn schema tax — this is the level the benchmark measured.
❯ reify init --write-agent-instructions
Appends a short block to AGENTS.md or CLAUDE.md. For tools
that read a different file — .cursorrules, CONVENTIONS.md,
.windsurfrules, .clinerules/ — paste the same lines.
Full instructions.
reify