🏛 Agora

English | 中文 | 日本語

Multi-agent AI that discusses, decides, and executes your tasks.

Open-source AI system where multiple agents debate your problem from different perspectives, then actually build the solution.

Why Agora?

• 🏛 Council, not chatbot — Multiple agents discuss your problem from different angles before acting.
• 🔧 Discussion → Execution — Agents don't stop at advice. They can write files, run commands, and implement the plan.
• 🧠 Self-improving — Discussions and executions are distilled into reusable skills over time.
• ⚙️ Customizable — Define your own agents, prompts, and models in YAML.
• 🔌 Model agnostic — OpenAI, Azure OpenAI, Claude CLI, Gemini CLI, Kiro CLI, and OpenAI-compatible APIs.
• 🐳 Self-hosted — Run it with Docker and keep control of your data.

Agora vs. other tools

	Agora	ChatGPT	AutoGPT	LangChain
Multi-agent debate before action	✅	❌	❌	DIY
Learns from discussion disagreements	✅	❌	❌	❌
Execution skills learned from use	✅	❌	⚠️	❌
Human-in-the-loop approval	✅	❌	⚠️	DIY
Self-hosted & open source	✅	❌	✅	✅

Agora builds on ideas from DeerFlow (sandbox + memory) and Hermes Agent by Nous Research (self-improving skills). Its original contribution is the council discussion model, which lets skills be learned not only from execution but from how multiple agents disagree and resolve.

What it looks like

You: "Add caching to my Go service (QPS ~5000)"

scout       → compares Redis vs Memcached
architect   → designs L1 + L2 cache strategy
critic      → points out consistency risks
synthesizer → generates action items

Execute? → yes

executor → writes code and runs commands

Quick Start

Docker

git clone https://github.com/wilbur-labs/Agora.git
cd Agora
cp .env.example .env  # edit .env and add your API key
docker compose up -d

Local

git clone https://github.com/wilbur-labs/Agora.git
cd Agora
cp .env.example .env  # edit .env and add your API key
make install
make dev

Configuration Example

models:
  gpt4o:
    provider: azure-openai
    api_key: ${AZURE_OPENAI_API_KEY}
    base_url: ${AZURE_OPENAI_BASE_URL}
    deployment: gpt-4o-0513

council:
  default_agents: [scout, architect, critic]
  model: gpt4o
  executor_model: gpt4o
  concurrent: false

How It Works

User Input
  → Moderator routes: QUICK / DISCUSS / EXECUTE / CLARIFY
    → QUICK: Single agent answers directly
    → DISCUSS:
        Scout → Architect → Critic → Synthesizer
        → User confirms action items
        → Executor runs tool-calling loop
        → Learn discussion + execution skills
    → EXECUTE:
        → Executor runs tool-calling loop directly
        → Learn execution skill

Council Agents

Agent	Role
Moderator	Routes requests
Scout	Research and evidence gathering
Architect	System design and solution planning
Critic	Review and challenge assumptions
Sentinel	Security review
Synthesizer	Summarizes decisions and action items
Executor	Executes with tools

Built-in Tools

Tool	Description
read_file	Read file contents
write_file	Create or overwrite files
patch_file	Update specific file content
list_dir	List directory contents
shell	Execute shell commands

Self-Learning

Agora learns from every interaction:

• Discussion skills — captures decision patterns and useful perspectives
• Execution skills — captures step-by-step implementation knowledge
• Memory — stores reusable user and project context
• Success tracking — records what works and what fails

CLI Commands

Command	Description
/ask <question>	Quick answer
/exec <task>	Direct execution
/agents	List council agents
/skills	List learned skills
/memory	View memory
/profile	View/set user profile
/reset	Clear conversation context
/quit	Exit

API

curl -N -X POST http://localhost:8000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Design a CI/CD pipeline for my Go project"}'

Testing

make test
make test-all

Roadmap

• Multi-agent discussion
• Tool-calling execution
• Self-learning skills
• Docker sandbox
• Multiple model backends
• Web UI
• Human-in-the-loop confirmation
• MCP server extensions
• Skill marketplace

Philosophy

In ancient Athens, the Agora was where people gathered to discuss, debate, and decide. Agora brings this idea to AI: not one model doing everything, but multiple perspectives collaborating before taking action.

License

MIT

Acknowledgments

• DeerFlow — inspiration for sandbox execution, memory systems, and orchestration
• Hermes Agent by Nous Research — inspiration for self-improving skills and persistent memory

Contact

• 📧 wilbur.ai.dev@gmail.com
• 🐛 GitHub Issues