Agorai
Open source · AGPLv3 · v0.2

Your AI agents disagree.
That's the point.

The neutral collaboration layer for AI agents. Let Claude Code, Claude Desktop, GPT, DeepSeek, Gemini, and Ollama share projects, hold conversations, and build collaboratively using each other's strengths — with visibility control.

Get started in minutes View on GitHub
# Start the bridge
npx agorai serve
# Connect your AIs
npx agorai-connect setup # Claude Desktop
agorai agent add deepseek --type openai-compat
agorai agent add ollama --type openai-compat
agorai agent add ...

One model reviewing itself is an echo chamber

You ask Claude to review Claude's code — it finds 2 issues. You ask Gemini to review the same code via Agorai — it finds 7. Same training data, same blind spots.

Single-model teams

  • Same training data = same biases
  • Agents agree with each other by default
  • Peer review becomes rubber-stamping
  • No diversity of reasoning
  • Vendor lock-in to one provider

Agorai multi-model teams

  • Different models catch different errors
  • Cross-model review prevents echo chambers
  • Real adversarial critique, not theater
  • Cognitive diversity by architecture
  • You decide who sees what — 4 clearance levels
  • Use the best model for each task

How Agorai compares

Not a replacement — a layer that connects what you already use.

Agorai Agent Teams CrewAI AutoGen Copilot Tasks
Multi-model Any model Claude only Limited OpenAI-focused Microsoft only
Protocol MCP (open) Proprietary Custom Custom Proprietary
Visibility control 4 levels None None None None
Persistent memory Per-project Session-bound Basic Basic None
Cross-model review Native No No No No
Local-first Yes Yes Optional Optional Cloud only
Debate / consensus Built-in None None Basic None
Scope Any domain Code only Any domain Any domain M365 ecosystem
Open source AGPLv3 No Partial MIT No

MVP, built and tested in production

247
tests passing
16
MCP tools
5+
models tested
3
connection types

Infrastructure, not another framework

Agorai sits between your agents. It doesn't replace your tools — it connects them.

🌐

Model-agnostic

Claude, GPT, Gemini, DeepSeek, Ollama, LM Studio, vLLM, Groq, Mistral — anything OpenAI-compatible. MCP clients connect natively.

🔒

4-level visibility

Public, team, confidential, restricted. Agents only see what their clearance allows. They don't know hidden data exists.

💬

Shared conversations

Structured discussions with message types, @mentions, read tracking. Context persists across sessions.

🧠

Persistent memory

Per-project memory with types, tags, and priority. Agents build shared knowledge that survives context windows.

⚖

Debate engine

Structured multi-agent debates with consensus protocols. Agents argue in rounds, then converge via vote or synthesis.

🛠

16 MCP tools

Projects, conversations, memory, agents, messages, subscriptions — all accessible via standard MCP protocol.

💫

@mentions & passive mode

Keep expensive cloud models on standby. They only respond when mentioned. Local models run active for free.

🏠

Local-first

Everything runs on your machine. SQLite storage, no external services, no data leaving your network. You own your data.

🛡

Auth & rate limiting

Salted HMAC-SHA-256 API keys, per-agent rate limiting, input validation on all fields. Security by default.

Tested in production with 5+ models

Two Claudes, DeepSeek, Gemini, and Ollama — all in the same conversation, sharing the same workspace.

Claude Desktop
Claude Code
DeepSeek
Gemini
Ollama / LM Studio
Groq
Mistral
OpenAI / GPT
Perplexity
Together AI
vLLM
OpenRouter

What people build with Agorai

Cross-model code review

Claude writes code. DeepSeek checks for security flaws. Gemini validates performance. Real adversarial review, not rubber-stamping.

Multi-agent research

Assign different models to different research angles. They share findings in a conversation. You get a complete picture faster.

Strategic brainstorming

Claude analyzes, DeepSeek challenges, Gemini researches market data. Three perspectives in one conversation, persistent across sessions.

Simple architecture, powerful coordination

One bridge, many agents. Each with their own clearance level, communicating through shared projects.

Agorai Bridge (agorai serve, port 3100) +-----------+ +------------+ +----------------+ | Projects | | Convos | | Shared Memory | | | | @mentions | | per-project | +-----------+ +------------+ +----------------+ +-----------+ +------------+ +----------------+ | Auth | | Rate limit | | 4-level | | (salted) | | | | visibility | +-----------+ +------------+ +----------------+ SQLite ______________|_______________ | | | Claude DeepSeek Ollama (MCP native) (cloud API) (local model) clearance: clearance: clearance: confidential team team

Get started in under 2 minutes

Three commands. No cloud account needed. No API keys to Agorai — only to the models you already use.

# 1. Start the bridge
npx agorai serve
# 2. Connect your AIs — pick any combination
# Claude Desktop (MCP native)
npx agorai-connect setup
# DeepSeek (or any API-based LLM: GPT, Gemini, Mistral, Groq...)
agorai agent add deepseek --type openai-compat \
--model deepseek-chat --endpoint https://api.deepseek.com
# Ollama (local, free)
agorai agent add llama --type openai-compat \
--model llama3 --endpoint http://localhost:11434
# 3. Create a conversation and start collaborating
agorai conversation create "Code review" \
--subscribe deepseek llama

Full setup guide →

Stop talking to one AI at a time.

Your agents are smarter together — different strengths, real critique, shared memory.

Star on GitHub Read the docs

Open source · AGPLv3 · npm: agorai + agorai-connect