autopilot-agent 1.0.3

Open source terminal-based AI coding agent built from scratch

Autopilot — AI Coding Agent (Built from First Principles)

Autopilot is a terminal-based AI agent that can understand entire codebases, execute complex multi-step tasks, manage long-running workflows, and operate autonomously within controlled limits.

It is built entirely in Python without using frameworks like LangChain, focusing on control, transparency, and reliability.

---

System Architecture

---

Overview

This project implements a complete agent system from scratch, including:

• Custom agent loop
• Tool calling system
• Context management (compaction + pruning)
• Session and checkpoint persistence
• Sub-agents for complex workflows
• MCP-based extensibility
• Safety and approval controls

The system is designed to run continuously until tasks are completed, while maintaining correctness and avoiding uncontrolled behavior.

---

Features

Core Functionality

• Interactive CLI mode for continuous workflows
• Single-run mode for automation tasks
• Streaming and non-streaming LLM responses
• Multi-turn reasoning with tool execution
• Retry mechanism with exponential backoff
• Configurable model, temperature, and behavior

---

Agent Loop

A fully custom-built execution loop:

• Interprets user intent
• Breaks tasks into steps
• Selects tools
• Executes actions
• Updates context dynamically

Runs until the task is completed or safely terminated.

---

Tool System

The agent interacts with the environment strictly through tools.

File Operations

• read_file, write_file, edit_file

Search & Navigation

• list_dir, glob, grep

Execution

• shell

Web

• web_search, web_fetch

Planning & Memory

• todos, memory

---

Context Management

Built for large codebases + long sessions:

• Compaction → compress past interactions
• Pruning → remove irrelevant tool outputs
• Token usage tracking
• Maintains continuity without overflow

---

Safety and Approval System

• Approval policies: auto, on-request, never, yolo
• Risk-aware execution
• Path validation
• Controlled shell access
• User confirmation for critical actions

---

Session & Checkpoints

• Save and resume sessions
• Create checkpoints mid-task
• Restore any previous state
• Persistent execution continuity

---

Subagents

• Spawn specialized agents for complex workflows
• Isolated execution contexts
• Useful for large refactors, audits, deep analysis

---

MCP Integration

• Connect external tools dynamically
• Supports stdio + HTTP/SSE
• Extend capabilities without changing core

---

Loop Detection

• Detects stuck/repeating behavior
• Prevents infinite loops
• Self-corrects execution

---

Terminal UI

Built using Rich + prompt_toolkit:

• Streaming responses
• Structured tool output
• Interactive commands

Currently Supported Commands

/help /config /model /tools /mcp /stats /save /resume /checkpoint /restore /history /exit

---

Tech Stack

• Pure Python without Langchain (completely built from scratch!)

---

High-Impact Use Cases

1. Full Codebase Refactoring

“Migrate this entire project from JavaScript to TypeScript and fix all type errors”

The agent:

• understands project structure
• updates configs + files
• fixes imports and types
• ensures consistency across the codebase

---

2. End-to-End Feature Implementation

“Add authentication with JWT + refresh tokens”

The agent:

• analyzes architecture
• creates middleware & routes
• updates database logic
• wires everything together

---

3. Deep Debugging & Root Cause Analysis

“Find why this API randomly fails in production”

The agent:

• traces execution paths
• inspects related modules
• identifies hidden issues
• suggests and applies fixes

---

Setup

Install

pip install autopilot-agent

Or install directly from PyPI: 👉 https://pypi.org/project/autopilot/

Configure

Create a .env file in your project root:

Option 1 — OpenRouter

OPENROUTER_API_KEY=your_api_key
BASE_URL=https://openrouter.ai/api/v1

Option 2 — Local Model (Ollama / LMStudio, etc.)

BASE_URL=http://localhost:11434/v1

Run

Just type autopilot in your terminal and see the magic unfold.