All Features
New

Build a support agent from your terminal

One command scans your codebase, detects your stack, and gives birth to an AI agent that answers customer questions with citations. Drafts stay marked until you review them. Run sync again after code changes, or wire scan and push into CI to refresh on deploy. Works with npx and bunx.

See what one command does

Run helpshelf scan and watch it map your entire project.

Terminal
$npx helpshelf scan
✔ Project: my-saas-app v2.1.0
✔ Stack: Next.js, React, Tailwind CSS, Stripe
✔ AI context: 1 file (8 sections)
✔ README: 12 sections
✔ API routes: 14 endpoints
✔ Dependencies: 28 packages
✔ Code: 45 components, 87 exports, 180 UI strings
✔ FAQs: 6 items generated
✅ Scan complete!
🧠 What we understood
Found user-facing content: "Pricing Plans",
"Sign Up Free", "Search your documents",
"Invite your team", and 94 more.
Your AI agent will answer questions like:
→ "What is my-saas-app and what does it do?"
→ "How do I get started?"
→ "What pricing plans are available?"

Your codebase, fully documented

The CLI does the heavy lifting so you don't have to

Auto-detect everything

Uses tree-sitter AST parsing to structurally understand your code — not just regex. Extracts framework, routes, env vars, and dependencies.

Function extraction

Parses every function signature with parameter types, return types, JSDoc comments, and async/export status via tree-sitter.

Import graph

Maps file-to-file relationships — imports, extends, implements. Understand how your codebase connects at a glance.

Test-as-docs

Mines describe/it blocks from test files and uses them as documentation. Your test suite becomes your knowledge base.

Parse your docs

Reads README files, markdown docs, and code comments. Extracts the structure and content your customers need.

Generate AI-ready snapshots

Creates a complete project snapshot — tech stack, dependencies, API reference, env vars, and auto-generated FAQs.

Capture screenshots

Point it at your running app and it captures screenshots of every page — including authenticated routes.

Publish in one command

Push your snapshot to HelpShelf. It lands as standard-tier knowledge your agent can answer from — and cite — right away.

Works with any project

Next.js, React, Vue, Express, Django, Rails — the CLI detects your stack and adapts its scanning automatically.

Three commands, done

From codebase scan to reviewable documentation

1

Initialize your project

npx helpshelf init

Creates a .helpshelf/ directory with your local configuration in one command.

2

Scan your codebase

npx helpshelf scan

Analyzes your project and shows what it understood — user-facing content, key features, and FAQ questions your AI agent will answer.

3

Publish to HelpShelf

npx helpshelf push

Sends your snapshot to HelpShelf. Your agent starts answering from it with citations, and AI-drafted content stays marked for your review.

What the scanner detects

UI text & labels
AI context files
API routes & endpoints
Environment variables
Dependencies & versions
README & markdown docs
Project structure
Framework & tech stack
Functions & signatures
Import relationships
Test descriptions

Keep docs in sync with CI/CD

Add the documented scan and push commands to your deploy pipeline to refresh HelpShelf after a successful code change. Review the generated content before publishing.

  • Runs in each pipeline where you add the commands
  • Pushes refreshed material for review
  • Works in CI environments that can run the CLI
  • Use --quiet flag for clean CI logs
.github/workflows/deploy.yml
- name: Update docs
  run: |
    npx helpshelf scan --quiet
    npx helpshelf push \
      --api-key ${{ secrets.HELPSHELF_KEY }}

Frequently asked questions

After scanning, the CLI shows a summary of what it learned about your product from the customer's perspective. It extracts headings, button labels, and content from your UI to understand what your app does — then generates FAQ questions your AI agent will answer. This is the "magic moment" that shows the scanner truly gets your product.

The CLI analyzes your UI text and labels, AI context files (like CLAUDE.md), README files, API routes, environment variables, dependencies, components, and documentation. It builds a complete snapshot that powers your AI support agent.

No — and we say so. Knowledge from the CLI scan lands as standard-tier content, and anything AI-drafted is marked generated until you review it. Content you write or import from your help desk is treated as curated (publisher-authored or imported) and ranks highest in answers. Trust tiers travel with every answer, so neither your customers nor external AI agents mistake a draft for verified docs.

No. The CLI only uploads the generated snapshot — metadata about your project like API routes, env var names (not values), dependency list, and parsed documentation. Your actual source code never leaves your machine.

The CLI auto-detects Next.js, React, Vue, Express, Fastify, Django, Rails, Laravel, and more. It works with any project that has a package.json, requirements.txt, Gemfile, or similar manifest.

Yes! Just use bunx instead of npx. For example: bunx helpshelf init, bunx helpshelf scan, bunx helpshelf push. Everything works the same way.

Yes. Add helpshelf scan && helpshelf push to the deployments where you want a refresh, and use --quiet for clean CI logs. A successful push uploads reviewable material; review and publish it in HelpShelf.

Tree-sitter is a parser used by VS Code, GitHub, and other tools for structural code understanding. HelpShelf uses tree-sitter via WebAssembly to extract function signatures, import graphs, and test descriptions from your code — no native compilation needed. If tree-sitter is unavailable, the CLI falls back to regex-based extraction.

Run helpshelf screenshot -u http://localhost:3000 and it captures every page of your running app using a headless browser. You can pass login credentials for authenticated pages. Screenshots are saved locally and can be pushed alongside your docs.

Ready to build your agent?

Install the CLI, scan locally, and review what it found before publishing.