AutoGPT/classic

AutoGPT Classic

AutoGPT Classic was an experimental project to demonstrate autonomous GPT-4 operation. It was designed to make GPT-4 independently operate and chain together tasks to achieve more complex goals.

Project Status

This project is unsupported, and dependencies will not be updated. It was an experiment that has concluded its initial research phase. If you want to use AutoGPT, you should use the AutoGPT Platform.

For those interested in autonomous AI agents, we recommend exploring more actively maintained alternatives or referring to this codebase for educational purposes only.

Overview

AutoGPT Classic was one of the first implementations of autonomous AI agents - AI systems that can independently:

• Break down complex goals into smaller tasks
• Execute those tasks using available tools and APIs
• Learn from the results and adjust its approach
• Chain multiple actions together to achieve an objective

Structure

classic/
├── pyproject.toml          # Single consolidated Poetry project
├── poetry.lock             # Single lock file
├── forge/                  # Core autonomous agent framework
├── original_autogpt/       # Original implementation
├── direct_benchmark/       # Benchmark harness
└── benchmark/              # Challenge definitions (data)

Getting Started

Prerequisites

• Python 3.12+
• Poetry

Installation

# Clone the repository
git clone https://github.com/Significant-Gravitas/AutoGPT.git
cd classic

# Install everything
poetry install

Configuration

Configuration uses a layered system:

1. Environment variables (.env file)
2. Workspace settings (.autogpt/autogpt.yaml)
3. Agent settings (.autogpt/agents/{id}/permissions.yaml)

Copy the example environment file and add your API keys:

cp .env.example .env

Key environment variables:

# Required
OPENAI_API_KEY=sk-...

# Optional LLM settings
SMART_LLM=gpt-4o                    # Model for complex reasoning
FAST_LLM=gpt-4o-mini                # Model for simple tasks

# Optional search providers
TAVILY_API_KEY=tvly-...
SERPER_API_KEY=...

# Optional infrastructure
LOG_LEVEL=DEBUG
PORT=8000
FILE_STORAGE_BACKEND=local          # local, s3, or gcs

Running

All commands run from the classic/ directory:

# Run forge agent
poetry run python -m forge

# Run original autogpt server
poetry run serve --debug

# Run autogpt CLI
poetry run autogpt

Agents run on http://localhost:8000 by default.

Benchmarking

poetry run direct-benchmark run

Testing

poetry run pytest                        # All tests
poetry run pytest forge/tests/           # Forge tests only
poetry run pytest original_autogpt/tests/ # AutoGPT tests only

Workspaces

Agents operate within a workspace directory that contains all agent data and files:

{workspace}/
├── .autogpt/
│   ├── autogpt.yaml              # Workspace-level permissions
│   ├── ap_server.db              # Agent Protocol database (server mode)
│   └── agents/
│       └── AutoGPT-{agent_id}/
│           ├── state.json        # Agent profile, directives, history
│           ├── permissions.yaml  # Agent-specific permissions
│           └── workspace/        # Agent's sandboxed working directory

• The workspace defaults to the current working directory
• Multiple agents can coexist in the same workspace
• Agent file access is sandboxed to their workspace/ subdirectory
• State persists across sessions via state.json

Permissions

AutoGPT uses a layered permission system with pattern matching:

Permission Files

File	Scope	Location
autogpt.yaml	All agents in workspace	.autogpt/autogpt.yaml
permissions.yaml	Single agent	.autogpt/agents/{id}/permissions.yaml

Permission Format

allow:
  - read_file({workspace}/**)     # Read any file in workspace
  - write_to_file({workspace}/**) # Write any file in workspace
  - web_search(*)                 # All web searches

deny:
  - read_file(**.env)             # Block .env files
  - execute_shell(sudo:*)         # Block sudo commands

Check Order (First Match Wins)

1. Agent deny → Block
2. Workspace deny → Block
3. Agent allow → Allow
4. Workspace allow → Allow
5. Prompt user → Interactive approval

Interactive Approval

When prompted, users can approve commands with different scopes:

• Once - Allow this one time only
• Agent - Always allow for this agent
• Workspace - Always allow for all agents
• Deny - Block this command

Default Security

Denied by default:

• Sensitive files (.env, .key, .pem)
• Destructive commands (rm -rf, sudo)
• Operations outside the workspace

Security Notice

This codebase has known vulnerabilities and issues with its dependencies. It will not be updated to new dependencies. Use for educational purposes only.

License

This project segment is licensed under the MIT License - see the LICENSE file for details.

Documentation

Please refer to the documentation for more detailed information about the project's architecture and concepts.