TalkingAgents Bedrock

A CLI chat agent powered by GPT-OSS 120B on AWS Bedrock with semantic tool discovery via Titan Text Embeddings V2.

Instead of sending all tool definitions in every request, the agent uses a gateway pattern: the LLM only sees a single discover_tools function and searches for relevant tools by describing what it needs. An embedding-based similarity search returns the top matches, which are then injected into the conversation dynamically.

Architecture

User message
    │
    ▼
LLM receives toolConfig = [discover_tools]
    │
    ▼
LLM calls: discover_tools({ description: "run a shell command" })
    │
    ▼
Gateway: embed query → cosine similarity vs tool embeddings → top-6
    │
    ▼
toolConfig updated: [discover_tools + matched tools]
    │
    ▼
LLM calls the appropriate tool → executes → responds

Tools

Tool	Description
get_datetime	Current date, time and timezone
calculate	Math expression evaluator (supports functions, exponentiation, parentheses)
generate_qr_code	Generates a QR code image URL
query_files	Reads files matching a regex and answers questions about them using AI
edit_file	Creates or modifies files via AI-generated search/replace diffs
shell	Executes shell commands in a specified directory
system_info	Machine details: OS, CPU, RAM, disk, uptime
web_search	Search the web via Brave Search API
web_fetch	Fetch a URL and answer questions about its content using AI

edit_file and shell require user confirmation before executing. You can approve, decline, or provide feedback explaining why (the agent will see your feedback and adjust).

Prerequisites

• Node.js 18+
• An AWS Bedrock account with access to:
  • openai.gpt-oss-120b-1:0 (or the model configured in config.js)
  • amazon.titan-embed-text-v2:0
• A bearer token for Bedrock API access

Setup

1. Clone the repository:

git clone https://github.com/your-username/TalkingAgentsBedrock.git
cd TalkingAgentsBedrock

2. Install dependencies:

npm install

3. Create a .env file with your credentials:

AWS_BEARER_TOKEN_BEDROCK=your_bearer_token_here
AWS_BEDROCK_REGION=us-west-2
BRAVE_API_KEY=your_brave_api_key_here

BRAVE_API_KEY is optional. Get one free at brave.com/search/api.

4. Run:

npm start

On first run, the gateway will compute embeddings for all tools (cached in .embeddings-cache.json for subsequent runs).

Configuration

Edit config.js to customize:

export default {
  mainModel: "openai.gpt-oss-120b-1:0",  // Chat + edit_file
  fastModel: "openai.gpt-oss-120b-1:0",  // query_files (swap for a cheaper model)
  maxMessages: 40,                         // Conversation sliding window
};

You can also change the terminal theme in main.js:

const THEME = "light"; // "dark" or "light"

Project Structure

├── main.js          # CLI chat loop and display
├── agent.js         # Agent class: streaming, tool loop, sliding window
├── tools.js         # Tool definitions and implementations
├── gateway.js       # discover_tools meta-tool factory
├── embeddings.js    # Titan V2 embedding service with SHA-256 cache
├── config.js        # Model and conversation settings
├── package.json
├── .env             # Credentials (not committed)
└── .gitignore

Example

You: What time is it and how much RAM does this machine have?

  [10:32:15.421] calling discover_tools({"description":"get current date and time"})
  [10:32:16.003] result  discover_tools -> {"discovered_tools":[{"name":"get_datetime",...}]}
  [10:32:16.450] calling get_datetime({})
  [10:32:16.451] result  get_datetime -> {"iso":"2025-06-15T10:32:16.451Z",...}
  [10:32:16.800] calling discover_tools({"description":"system memory RAM information"})
  [10:32:17.200] result  discover_tools -> {"discovered_tools":[{"name":"system_info",...}]}
  [10:32:17.500] calling system_info({})
  [10:32:17.502] result  system_info -> {"ram":{"total":"16.0 GB","free":"8.2 GB",...},...}

Assistant: It's 10:32 AM (UTC). Your machine has 16 GB of RAM total with 8.2 GB free.

Adding New Tools

Add your tool definition to the toolDefinitions array in tools.js and its implementation to the toolImplementations object. The embedding cache will automatically detect the new tool on next startup and compute its embedding.

License

MIT

---

Made with ❤️ by Juan Pablo Gramajo