Use Cases

What Codehere does well today — and where it has limits.

Features below work in v0.1.1 (current npm release). In-progress items are labelled clearly.

What Works Well Today

Understand an Unfamiliar Codebase

You inherited a codebase with no documentation. Index it, then ask natural language questions to quickly find where things are and how they work.

Try it:

codehere ask "where is authentication handled?"

What it does:

  • Builds semantic embeddings locally (SQLite)
  • Returns relevant files and code snippets
  • Works best with TypeScript/JavaScript codebases
  • Per-repository isolation — no cross-project leakage

Limitation: Best for repos under 100K lines. Large repos may be slow to index.

Safe AI-Assisted Code Editing

Fix files with AI assistance. Every instruction is scanned for prompt injection first. Every generated change passes SAST before touching your files.

Try it:

codehere fix src/api.ts "add input validation"

What it does:

  • Prompt injection scan on every instruction
  • SAST scanning on every AI-generated edit
  • Automatic backup before changes
  • Rollback if something goes wrong

Limitation: Single-file editing only. Multi-file editing is on the roadmap.

Plan Complex Changes Before Doing Them

Break down a big task into steps before writing any code. See which files will be affected, what the dependencies are, and what the risks are.

Try it:

codehere plan "refactor auth to use JWT tokens"

What it does:

  • Step-by-step task breakdown
  • Identifies affected files and dependencies
  • Risk assessment for each step
  • Uses codebase context for realistic plans

Limitation: Plans are AI-generated suggestions, not guaranteed-correct blueprints. Always review.

Security-Conscious Development

Every AI-generated code change passes through a safety pipeline. Catches SQL injection, hardcoded secrets, unsafe patterns — before they reach your codebase.

Try it:

codehere fix src/db.ts "add caching"

What it does:

  • Prompt injection detection on the instruction
  • AI SAST scanner checks generated code
  • Detects hardcoded secrets and passwords
  • Catches unsafe patterns: eval, exec, path traversal, ReDoS

Limitation: SAST is pattern-based. Not a replacement for a professional security audit.

Privacy-First AI Assistance

Run completely offline with Ollama. Or use cloud providers — but your code index always stays local. No telemetry. No data collection.

Try it:

export CODEHERE_AI_PROVIDER=local codehere ask "explain this function"

What it does:

  • Local SQLite storage for all data
  • Works fully offline with Ollama
  • No telemetry or analytics — ever
  • Per-repo data isolation

Limitation: Local models (Ollama) require a machine with at least 8GB RAM. Quality varies by model.

Multi-Step Orchestration

Coordinate multi-step changes across your codebase. Codehere breaks down the task, executes steps in order, and validates along the way.

Try it:

codehere orchestrate "add error handling to all API routes"

What it does:

  • Multi-agent coordination
  • Step-by-step execution with validation
  • Uses codebase context for accurate changes
  • Review gates at each step

Limitation: Complex orchestrations with many files can be slow and may need manual intervention.

What We're Building Next

These are built or in progress — targeting v0.2.0. Not yet published to npm. See the full roadmap for details.

Safety Benchmark (50 Patterns)

Built, not published

A public, reproducible benchmark: 50 attack patterns across 5 categories. Built and runnable from source. Targeting inclusion in v0.2.0 npm release.

Prompt Injection Detection

Built, not published

Every instruction scanned for adversarial patterns before it reaches the AI. Built into the safety pipeline. Targeting v0.2.0.

VS Code Extension

Untested alpha

Alpha code exists but is untested with real users. Not published to the VS Code Marketplace yet. Do not rely on it.

Multi-File Editing

In progress

Coordinate changes across multiple files in a single operation with dependency awareness.

Known Limitations

We believe in honesty. Here's what Codehere doesn't do well yet:

Large repos (100K+ lines) are slow to index
Best support for TypeScript/JavaScript — other languages have limited support
Multi-file editing is in progress, not yet released
Conversation history doesn't persist across sessions yet
AI quality depends on your chosen provider and model
Orchestration can be slow for complex multi-step tasks
Safety benchmark is pattern-based — AI-powered mode requires an API key
Embeddings should be rebuilt after major code changes

Try it yourself

Free, open source, works offline. Install and start asking questions in under 2 minutes.