plan_change
Build an evidence-backed edit plan with primary context, impact candidates, validation targets, and boundary policies — before touching a single file.
Connect Open Kioku as a local MCP server so Claude Code shows its evidence before it edits — files, streams, confidence, caveats — declares an edit boundary in a plan, and has its diff verified against that plan. Nothing leaves your machine.
Install, index, connect, prompt. Indexing time depends on repo size; Open Kioku commands run locally on your machine.
Install the CLI globally. The npm wrapper pulls the native binary for macOS, Linux, or Windows automatically.
$ npm install -g open-kiokuOther channels: cargo install open-kioku-cli or cargo binstall open-kioku-cli from crates.io; binaries with checksums, an SBOM, and provenance on GitHub releases.
Navigate to your project directory and run the indexer. Open Kioku scans files, extracts symbols, builds a dependency graph, and creates local SQLite + Tantivy search indexes under .ok/.
$ ok index .Shortcut: one command indexes the repository, writes a repository-scoped .mcp.json plus managed guidance, and checks that the local server answers an MCP initialize request (run without --apply to preview; nothing is written).
$ ok setup agent claude --repo . --applyOr generate the MCP server entry for Claude Code manually and paste it into your Claude Code MCP settings:
$ ok mcp install claude --repo .{
"mcpServers": {
"open-kioku": {
"command": "ok",
"args": [
"mcp",
"serve",
"--repo",
"/absolute/path/to/repo",
"--read-only"
]
}
}
}Add this prompt to your CLAUDE.md file or system instructions. It tells Claude Code to gather evidence from Open Kioku before every edit and verify changes afterward.
Use Open Kioku before editing. Check repo_status, search_code, get_definition,
get_references, impact_analysis, and find_tests_for_change. Build a plan with
plan_change first, then edit, and verify after the edit with verify_change.Every Open Kioku tool runs locally against your indexed repo. No hosted index or embeddings service required, read-only by default.
What to expect: on a Java repository of about ten thousand files the right file is in the top five about half the time and in the context pack about two thirds of the time; on a TypeScript repository of about nine hundred files, in the pack nearly nine in ten and in the top five about four in five. That is the retrieval floor before exact lookups and the plan → verify loop; it is re-measured nightly against frozen baselines. Numbers and method.
Build an evidence-backed edit plan with primary context, impact candidates, validation targets, and boundary policies — before touching a single file.
After editing, verify the change stayed within the plan's boundaries and didn't break expected contracts.
BM25-ranked code search returning snippets with file paths, line ranges, and confidence scores.
Jump to the definition of any symbol — functions, classes, types — from the local symbol index.
Find every reference to a symbol across the codebase using indexed graph edges, not text matching.
Trace which modules, callers, and dependents a proposed change would affect before it happens.
Identify which tests cover the files and symbols you're about to change, so the agent runs the right ones.
Check index health, file counts, symbol totals, and last-indexed timestamps at a glance.
Assemble a focused context window of relevant files, symbols, and tests for a specific task.