Code Agent for Jira Tasks

Turn Jira tickets into
precise code context

Semantic search over your repository. Finds exact functions, classes, and dependencies needed for any task — then packs them into AI-ready XML.

Get Started

~3s Search time

AST Block-level index

384d Vector embeddings

bash — crang7

$ crang7 index ./project

📦 repomix-output.xml generated

🔍 Found 233 files, 220 AST blocks

💾 LanceDB: 220 blocks, 125 files indexed

✅ Index ready (commit: a1b2c3d)

$ crang7 search "add Google OAuth2 auth"

vector blocks: 40, files: 10

after reranking: 15

+ 2 related files via import graph

✅ XML: repo-context-1716123456789.xml

Files: 5 (primary: 3, related: 2)

Pipeline

From raw repository to AI-ready context in 4 stages

Ingest

Respects .gitignore. Lazy repomix XML. Collects only changed files on re-index.

Parse

TypeScript AST extractor. Functions, classes, interfaces, types. Smart chunking for large functions.

Embed

all-MiniLM-L6-v2 (384d) via transformers.js. SHA256 cache. Cross-encoder reranking with ms-marco.

Output

Repomix-compatible XML. Import graph expansion. [RELATED] markers. Truncation where code is irrelevant.

Why crang7

Built for teams that ship fast with AI assistance

Incremental Indexing

Only re-indexes changed files via git diff. Full rebuild takes seconds, not minutes. SHA256 embedding cache persists across runs.

AST-Level Precision

Not file-level — function-level. Finds the exact `authenticate()` method, not the entire 2000-line auth module. Chunks large functions by control flow.

Cross-File Graph

Resolves imports and call references automatically. Pulls in type definitions, utilities, and middleware your task depends on — even if they don't match the query semantically.

Local & Private

Everything runs on your machine. No API keys, no cloud uploads, no data leakage. transformers.js + LanceDB are fully local.

Repomix Native

Output follows the standard repomix XML format. Kimi, Claude, GPT — all parse it natively. No custom prompts needed.

Scope Filtering

Index and search within specific directories. `--scope src/auth` keeps focus tight when your codebase grows.

~45MB Binary

Includes Node.js runtime, transformer models, and LanceDB. Download and run.

Cross-Platform

Linux, macOS, Windows. Same CLI, same behavior everywhere.

What Kimi receives

Not walls of code. Only the exact blocks needed for your Jira task, with context preserved.

Relevant functions with signatures
Imports for dependency understanding
[RELATED] files pulled via graph
Truncation markers for skipped code
Chunk markers for large functions
Semantic relevance scores per block

                    repo-context-1716123456789.xml
                
                    <file path=<"src/auth/providers.ts">

                    // ===== imports =====

                    import { OAuth2Client } from 'google-auth-library';

                    import { User } from '../models/user';

                    import { signJwt } from '../utils/jwt';

                    // === function: createGoogleStrategy (combined: 0.891) ===

                    export function createGoogleStrategy(config: OAuthConfig) {

                     return new OAuth2Client(config.clientId, ...);

                    }

                    // ... [340 chars omitted] ...

                    // === class: GoogleAuthProvider (combined: 0.854) ===

                    export class GoogleAuthProvider { ... }

                    </file>

                    <file path=<"src/utils/jwt.ts">

                    // === function: signJwt [RELATED] (dist: 0.234) ===

                    export function signJwt(payload, expiresIn = '7d') { ... }

                    </file>

Turn Jira tickets into precise code context