opentf 0.1.0

Open Task Force - Autonomous multi-agent orchestration platform

OpenTF

Autonomous AI coding agent for your terminal. Type plain English, agents handle the rest.

OpenTF runs a pipeline of specialized AI agents -- security guardrails, context enrichment, planning, code generation, and quality scanning -- all orchestrated through a single conversation loop with a rich terminal UI.

Quick Install

curl -fsSL https://opentf.dev/install.sh | bash

Requires Python 3.12+. See Installation docs for detailed setup, updating, and uninstalling.

Usage

opentf

On first run, you'll be prompted to enter your Anthropic API key. You can also set it via environment variable:

export ANTHROPIC_API_KEY="sk-ant-..."
opentf

Commands

Command	Description
/plan	Create a structured execution plan
/janitor	Scan code for quality issues
/model	Switch between Claude models
/compact	Compress conversation history
/status	Show current session info
/cost	Show token usage and cost
/login	Set API key
/logout	Remove stored API key
/clear	Clear conversation
/exit	Quit OpenTF

How It Works

User prompt
  |
  v
SecurityAgent (blocks injections, path traversal, command injection)
  |
  v
ContextAgent (enriches with workspace info + memory retrieval)
  |
  v
MainAgent (code, files, research, data -- unified tool access)
  |
  v
Response (streamed to TUI with syntax highlighting)

MainAgent handles everything through a single conversation loop with file tools, data tools, and shell access. Optional specialists can be invoked:

• PlannerAgent -- decomposes tasks into phased execution plans
• JanitorAgent -- scans for dead code, unused imports, bad naming, silent failures
• SkillBuilderAgent -- dynamically creates new specialist agents at runtime

All agents communicate through a protocol-based message bus with immutable task snapshots for safe retries.

Development

git clone https://github.com/itscool2b/opentf.git
cd opentf
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
opentf

Run tests:

pytest

Project Structure

src/opentf/
  agents/          # Agent system (base, registry, guardrails, specialists)
  auth/            # API key management
  cli/             # Textual TUI (app, widgets, theme)
  core/            # Engine, orchestrator, message bus, tool loop
  llm/             # Anthropic client wrapper
  memory/          # ChromaDB vector store + BM25 hybrid retrieval
  models/          # Pydantic models (task, message, plan, context)
  tools/           # File tools, data tools
config/            # Default YAML configuration
docs/              # Architecture and setup documentation
tests/             # Test suite

Key Design Decisions

• Single-agent loop over multi-agent routing -- simpler, fewer LLM calls, unified tool access
• Guardrails as gates -- security and context run pre-execution, not as validators
• Protocol-based messages -- deterministic type routing, no free-form text parsing
• Hybrid memory -- vector search (ChromaDB) + keyword search (BM25) via reciprocal rank fusion
• Async-first -- all core operations use async def

License

MIT