Open Kioku is a local index your coding agent reads before it edits. It shows its evidence — which files, which streams answered, what is missing — declares an edit boundary in a plan, and verifies the finished diff against that plan. Nothing leaves your machine.
Retrieval is scored on the production ok context path — the same builder behind ok plan and the MCP build_context_pack tool — on four repositories indexed at a fixed base commit. Each case is a later commit: the query is its subject line, the answer is the source files it changed. Both splits are gated nightly; the table shows holdout.
| Corpus | Holdout cases | R@5 | R@20 | MRR |
|---|---|---|---|---|
| Java, about 10k files | 113 | 0.566 | 0.708 | 0.516 |
| Go application, ~800 files | 84 | 0.690 | 0.821 | 0.539 |
| TypeScript, ~900 files | 167 | 0.826 | 0.880 | 0.683 |
| Python library, ~4k files | 199 | 0.663 | 0.754 | 0.543 |
bd5d06ab, confirmed by a second run of the same commit with identical scores, and are re-derived nightly against the frozen files in benchmarks/commit-derived/ by the scripts in the repository; the job fails when a watched metric drops more than 0.03 below its baseline, or when a routed task family of 34 or more cases, with unchanged membership, drops more than max(0.03, 2/n) below its own. Queries are commit subjects, not issue text, so these numbers are not comparable with published benchmarks. Method, corpora, both splits, and scripts ↗4.0.0 shipped on 2026-09-11. It changes the index format (run ok index after upgrading) and cuts the MCP surface to sixteen tools. Every figure below has its method in CHANGELOG.md or the linked doc, and names its commit and artifact where one exists; the retrieval figures are local-workstation A/B runs, not the nightly hosted baselines above.
Every advertised tool answers one question no other tool answers. Twenty-three names folded into the sixteen with nothing lost; thirteen moved to the CLI and still ship; one (structural_search) was removed because no structural matching existed, only the name. A retired name answers with where its capability went, not a bare refusal. Six descriptions that said what their names suggested now say what the implementation does.
regex_search now does regexIt advertised regular-expression matching and dispatched to ranked lexical search, so an agent that asked for a pattern got guesses with nothing in the response saying so. It now compiles the pattern once, evaluates it line by line over indexed text, reports how many files it scanned, and says when it stopped early. ok search <pattern> --regex makes the same call.
Opening a pre-4.0 index withholds relationship evidence and answers run ok index to rebuild. ok impact, ok plan, ok context and the MCP tools built on them refuse rather than answer from an empty graph, because an empty answer reads as “no such relationship exists”. Search and symbol lookup keep working meanwhile.
Per-language coverage in ok index, ok doctor, ok status and repo_status, with every omission attributed to a skip reason. An ingest rule had silently dropped 25 Java source files from one repository, and nothing surfaced it. Source files whose path contains “secret” are now indexed with their parser messages redacted; generated files are indexed, flagged, and ranked last unless the task names them.
Packs showed the right file but the wrong cut: selected units covered 3–22% of the lines the real commit changed. The top three files now widen to the enclosing symbol and adjacent chunks. Share of changed lines shown within 8k tokens: Java 0.216 → 0.248, Go 0.207 → 0.299, TypeScript 0.155 → 0.335, Python 0.130 → 0.203 (626 paired cases, no case worse, about three times the tokens; the Go corpus reads high because 50 of its cases are near-identical release bumps).
ChannelsUtils Tests reaches ChannelUtilsTests; a misspelt part the repository spells nowhere is corrected by one edit. No model, no network, no re-index. Neutral on commit-subject benchmarks by construction; on 259 perturbed queries, R@5 0.656 → 0.699 and MRR +0.036 (95% CI +0.015 to +0.062). Read that as an upper bound: the perturbations are the inverse of the lattice’s own hypothesis, and this measurement has no committed artifact.
The most recent end-to-end scale record validated the 3.1.0 release lineage (source 3959fdf) on a large Java repository — same protocol and same hardware as the 3.0.4 record; the before/after figures compare against main at c96f61a on the identical corpus. 4.0.0 changed the index storage format and has not been re-run on this corpus; these numbers describe 3.1.0.
main at c96f61a, same host as the 3.0.4 recordEvery claim on this page traces to a frozen corpus and a machine-readable artifact in the repository — including the numbers that are not good yet.
On the frozen 30-case fixture, none of the five no-gold tasks is presented as an answer: the no-gold false-positive rate is 0.0 (the CI ceiling is 0.25), and the advisory routed context-pack strategy in the same run returns none of them above Low. A low-confidence pack still lists candidates — it tells the caller not to trust them rather than returning nothing. Calibrated abstention is built and gated on held-out performance, and off by default until it earns activation.
A file that merely shares a name can never become a “proven dependency” or a “required test.” Structural claims carry typed proofs; heuristic candidates stay clearly labeled suggestions.
Every index builds into its own generation and publishes with one atomic switch — designed so a reader never sees a half-built index; a kill mid-build or a downgrade falls back to the last good generation.
The evidence below identifies the exact build lineage, publishes aggregate results, and keeps caveats beside the claims they qualify. The scale record is from the 3.1.0 lineage; 4.0.0’s own measured changes are listed above with their commits and methods.
Validated end to end on a large Java workload — and honest about the result. The adjacent artifacts are the source of record; the replay on the left shows output in the shape the CLI prints it.
The JSON records build source, host profile, configuration, workload totals, measurements, quality checks, and limitations — including what still isn’t perfect. Repository identity is withheld, so exact replay is not claimed.
The repeated structural run reproduced identical totals, the exact class lookup returned the correct class in 0.02s, and missing compiler-grade SCIP evidence stayed visible instead of being treated as present.
Open Kioku combines deterministic code intelligence with explainable retrieval signals. Exact repository facts do not get demoted by a fuzzy match.
Definitions, references, paths, and exact evidence stay authoritative. Heuristics can help retrieval without rewriting repository truth.
Ranked snippets, exact line ranges, impact, nearby tests, and caveats give the next agent step a bounded evidence set.
Optional semantic retrieval stays on-device and can use persistent approximate nearest-neighbor search while exact-flat remains the correctness oracle.
Co-change, churn, ownership, reviewer, provenance, renames, and similar-change signals are derived from bounded local Git history.
Validation candidates are selected from indexed tests and repository evidence before a change begins, not as an afterthought.
Plans define edit boundaries. Verification compares the actual changed files with those boundaries and reports evidence caveats.
A terminal recording against a 39-file fixture: ok setup agent indexes and wires the agent, ok context shows evidence and confidence, ok plan writes the boundary, then one edit inside it and one drive-by outside it. ok verify fails on the drive-by. Rendered from demo.tape; anyone with ok installed can re-record it.
Click through the same sequence an agent can use before, during, and after a repository change. The verify step is the attested dogfood run in verification-dogfood.json, produced by ok 2.4.0. The 12-second recording under the tour shows that step on its own.
![Terminal recording: the allowed files from plan.json, one edit inside that boundary, one edit outside it, then ok verify prints Verification: Fail, reports [out_of_boundary] go/shipping/carrier.go, and exits 1.](media/demo-verify.gif)
The default workflow is designed for local repository intelligence rather than a hosted code-index service.
Core search and evidence are stored under the repository’s local .ok/ data.
The MCP server provides evidence tools; source edits remain in the normal editor and agent workflow.
Command execution and model downloads are policy-gated. Default configuration includes network denial, and denial fails closed. Local semantic search needs no hosted embeddings API.
ok prove reports index and validation signals while intentionally omitting source snippets.
Install from the channel you already use, run one setup command, then ask for evidence on the task you actually care about.
npm install -g open-kiokuPulls the native binary for macOS, Linux, or Windows through platform packages.
cargo install open-kioku-cliOr cargo binstall open-kioku-cli for the prebuilt binary.
ok-{linux,macos,windows}-*Binaries with SHA256SUMS, an SBOM, PROVENANCE.json, and build-provenance attestations. Release trust ↗
claude_plugin.jsonPlugin manifest and .claude-plugin/ marketplace entry in the repository.
.cursor-plugin/ · .codex-plugin/Agent, rules, and skill for Cursor; plugin manifest for Codex.
glama.json · smithery.yamlServer manifests for the Glama and Smithery MCP directories.
ok setup agent claude|cursor --repo . --applyClaude Code · Cursor · Codex · Gemini CLI