@emisso/talk — Talk with your Software

An open-source SDK that adds an AI chatbox to your app where users can ask questions about your software. Powered by Claude Code running in Vercel Sandbox against your actual source code.

"How do I set up SSO?" "Where do I add a new user?" "How does billing work?"

Your users ask questions. Claude Haiku explores your codebase. They get accurate, up-to-date answers.

How it works

User asks a question
       ↓
Query classified (code_lookup, architecture, troubleshoot, etc.)
       ↓
Vercel Sandbox microVM created
       ↓
Your repo cloned into the sandbox
       ↓
Claude Code (Haiku) explores the code with Read, Grep, Glob, Bash
       ↓
Structured answer streamed back to the widget

No RAG. No embeddings. No vector database. Claude reads your actual code every time.

Quick Start

1. Install

pnpm add @emisso/talk-next @emisso/talk-react @ai-sdk/anthropic ai

2. Create the API route

// app/api/talk/route.ts
import { createTalkRouter } from '@emisso/talk-next'

export const { POST } = createTalkRouter({
  anthropicApiKey: process.env.ANTHROPIC_API_KEY!,
  repo: 'your-org/your-app',
  githubToken: process.env.GITHUB_TOKEN!,
})

3. Add the widget

// app/page.tsx
import { TalkWidget } from '@emisso/talk-react'

export default function Page() {
  return (
    <>
      <YourApp />
      <TalkWidget
        endpoint="/api/talk"
        appName="My App"
        primaryColor="#10b981"
      />
    </>
  )
}

That's it. Your users can now ask questions about your software.

Packages

Package	Description
@emisso/talk	Core engine — sandbox, Claude provider, query classifier, skills
@emisso/talk-react	React chat widget — floating chatbox with animations
@emisso/talk-next	Next.js adapter — createTalkRouter() for App Router

Configuration

createTalkRouter options

Option	Type	Default	Description
anthropicApiKey	string	required	Your Anthropic API key
repo	string	required	GitHub repo (owner/name)
githubToken	string	required	GitHub PAT for cloning
model	string	claude-haiku-4-5-20251001	Claude model ID
maxTurns	number	20	Max CLI agent turns per query
timeoutMs	number	120000	Execution timeout (ms)
vcpus	number	2	Sandbox vCPUs (1-8)
revision	string	undefined	Branch/tag/commit
systemPrompt	string	undefined	Custom system prompt
instructionFileContent	string	undefined	CLAUDE.md content
onRequest	function	undefined	Middleware (auth, rate limit)

<TalkWidget> props

Prop	Type	Default	Description
endpoint	string	required	API route URL
appName	string	"Assistant"	Name shown in header
position	string	"bottom-right"	bottom-right or bottom-left
primaryColor	string	"#10b981"	Accent color
greeting	string	"Ask me anything!"	Welcome message
placeholder	string	"Type a question..."	Input placeholder
launcherIcon	string	"chat"	chat, help, or question
onOpen	function	undefined	Called when widget opens
onClose	function	undefined	Called when widget closes

Programmatic API

const talkRef = useRef<TalkWidgetRef>(null)

talkRef.current?.open()
talkRef.current?.close()
talkRef.current?.toggle()

Query Categories

The engine classifies each question and selects the right exploration strategy:

Category	Description	Example
code_lookup	Find specific code	"Show me the auth middleware"
architecture	Explain system structure	"How does billing work?"
flow_trace	Trace execution paths	"What happens when a user logs in?"
capability_check	Does feature X exist?	"Do we support SSO?"
implementation_how	How to implement X	"How would I add webhooks?"
troubleshoot	Debug a problem	"Why is webhook delivery failing?"
general_product	Product usage	"What fields are required to create an event?"

Advanced Usage

Custom instruction file

Write a CLAUDE.md for your repo to give the agent context:

export const { POST } = createTalkRouter({
  // ...
  instructionFileContent: `
# My App

## Tech Stack
Next.js, Prisma, PostgreSQL, Stripe

## Key Patterns
- All API routes use middleware chain in src/middleware/
- Database models in src/db/schema/
- Business logic in src/services/
  `,
})

Auth middleware

export const { POST } = createTalkRouter({
  // ...
  onRequest: async (req) => {
    const token = req.headers.get('Authorization')
    if (!token || !isValidToken(token)) {
      return Response.json({ error: 'Unauthorized' }, { status: 401 })
    }
  },
})

Using the engine directly

import { navigate } from '@emisso/talk'

const result = await navigate(
  {
    anthropicApiKey: process.env.ANTHROPIC_API_KEY!,
    repo: 'owner/repo',
    githubToken: process.env.GITHUB_TOKEN!,
    model: 'claude-haiku-4-5-20251001',
  },
  { query: 'How does authentication work?' }
)

console.log(result.answer)
console.log(result.codeSnippets)
console.log(`Cost: $${result.costUsd}`)

Requirements

• Node.js >= 18
• Anthropic API key (get one at console.anthropic.com)
• GitHub personal access token (for repo cloning)
• Vercel deployment (for Vercel Sandbox microVMs)

Cost

Claude Haiku is very affordable:

• Input: $0.80 / million tokens
• Output: $4.00 / million tokens
• A typical question costs $0.01 - $0.05

License

MIT

Contributing

Built by Emisso. PRs welcome.