OpenAgent

OpenAgent is a Node.js-based CLI agent that communicates with an LLM using a structured JSON protocol. It safely executes shell commands with human-in-the-loop approval, live progress tracking, and output controls.

If you want the very slim, no extra fluff version, check this early commit: https://github.com/asynkron/OpenAgent/tree/9073a48225ddc2c2afc6f701af48d6ec4c04e477 - 400 LoC

Example chat

Plan ______________
✔ Step 1 - Review brain directory guidance
✔ Step 2 - Inspect relevant CLI modules for human-interaction flow
✔ Step 3 - Draft TODO plan for implementing --nohuman flag
▶ Step 4 - Implement --nohuman flag behavior in CLI
• Step 5 - Run syntax/tests and verify behavior
• Step 6 - Summarize changes and next steps

Command ______________
Shell: bash
Directory: /Users/rogerjohansson/git/asynkron/OpenAgent
Timeout: 60s
Auto-approved by CLI flag (--auto-approve)

Command Result ______________
Exit Code: 0
Runtime: 3ms
Status: COMPLETED

STDOUT ______________
  Edited src/agent/loop.js
  "use strict";

  /**
   * Implements the interactive agent loop that powers the CLI experience.
   *
   * Responsibilities:
   * - Maintain the conversation history with OpenAI responses.
   * - Render assistant output, prompt the human for approvals, and execute commands.
   * - Feed execution results back to the model to continue the workflow.
   *
   * Consumers:
   * - Library entry `src/lib/index.js` creates a configured loop via `createAgentLoop()` and exposes it as `agentLoop`.
   * - Integration tests re-export the same function to run mocked scenarios.
   */

  import chalk from 'chalk';

  import { SYSTEM_PROMPT } from '../config/systemPrompt.js';
  import { getOpenAIClient, MODEL } from '../openai/client.js';
  …



Sending request to AI
Received response from AI

Command ______________
Shell: bash
Directory: /Users/rogerjohansson/git/asynkron/OpenAgent
Timeout: 60s

  sed -n '200,300p' src/agent/loop.js


Auto-approved by allowlist (approved_commands.json)

Table of Contents

• Features
• Getting Started
• Configuration
• Usage
• JSON Protocol
• Templates & Shortcuts
• Safety
• License

Features

• Structured JSON Protocol for deterministic LLM↔agent communication
• Command Execution with Timeouts and interactive approval
• Plan Visualization showing task progress as a checklist
• Output Filtering via regex and tail limits
• Conversation History persisted across steps for context retention

Getting Started

1. Clone the repository:

   git clone <your-fork-url>
   cd openagent

2. Install dependencies:

   npm install

3. Copy the environment template and add your OpenAI API key:

   cp .env.example .env

Quality tooling

• npm run lint — runs ESLint with the shared configuration.
• npm run format — formats the repository with Prettier.
• npm run format:check — verifies committed files already match the Prettier style.

Configuration

Edit .env to include the required secrets:

OPENAI_API_KEY=sk-...
# Optional: override the default model (defaults to gpt-5-codex)
# OPENAI_MODEL=gpt-4.1-mini
# Optional: override the assumed model context window (tokens, defaults to 256000)
# OPENAI_CONTEXT_WINDOW=256000
# Optional: set reasoning effort for reasoning-enabled models (one of low, medium, high)
# OPENAI_REASONING_EFFORT=medium
# Optional: enforce a request timeout (milliseconds)
# OPENAI_TIMEOUT_MS=60000
# Optional: configure automatic retry attempts for transient API errors
# OPENAI_MAX_RETRIES=2
# Optional: point at a custom Responses API endpoint (must be the API root, not a completions URL)
# OPENAI_BASE_URL=https://api.openai.com/v1

The client validates that the configured model supports the Responses API and that any custom base URL targets the API root, so misconfigurations fail fast.

Usage

Start the CLI agent:

npm start

Interaction loop:

1. Type a task and press Enter
2. Review the agent’s explanation and plan
3. Inspect proposed commands (they won’t run without approval)
4. Press Enter to execute, or type skip
5. View results; repeat until the task completes

To exit, type exit or quit at any prompt.

JSON Protocol

LLM → Agent messages look like:

{
  "message": "status update",
  "plan": [{ "step": 1, "title": "Describe goal", "status": "running" }],
  "command": {
    "shell": "bash",
    "run": "ls -a",
    "cwd": ".",
    "timeout_sec": 60,
    "filter_regex": "error|warning",
    "tail_lines": 200
  }
}

Agent → LLM observations:

{
  "observation_for_llm": {
    "stdout": "output",
    "stderr": "",
    "exit_code": 0,
    "truncated": false
  },
  "observation_metadata": {
    "runtime_ms": 120,
    "killed": false,
    "timestamp": "2025-02-01T12:00:00Z"
  }
}

Templates & Shortcuts

• Command templates in templates/command-templates.json:
  • List: npx openagent templates list (or node ./bin/openagent.js templates list)
  • Show: npx openagent templates show <id>
  • Render: npx openagent templates render <id> '{"var":"value"}'
• Shortcuts in shortcuts/shortcuts.json:
  • List: npx openagent shortcuts list
  • Show: npx openagent shortcuts show <id>
  • Run (prints command): npx openagent shortcuts run <id>

Safety

• Every command has a configurable timeout
• Execution always requires explicit user confirmation
• Outputs can be limited or filtered to avoid flooding
• Errors are captured and surfaced cleanly

License

This project is released under the MIT License.