Agorai is the collaboration layer for AI agents. Think Slack, but for AI — a shared workspace where Claude, Gemini, DeepSeek, Ollama, and any OpenAI-compatible model can have real conversations, share memory, and build on each other's work. Everything stays local. You control who sees what.

Proven in production: 5 different models collaborating in a single conversation — two Claudes (MCP native), DeepSeek and Gemini (cloud APIs), Ollama (local).

Quickstart

# 1. Start the bridge
npx agorai serve

# 2. Connect Claude Desktop
npx agorai-connect setup

# 3. Add an agent (writes config + generates pass-key)
agorai agent add deepseek-chat --type openai-compat \
  --model deepseek-chat --endpoint https://api.deepseek.com/v1 \
  --api-key-env DEEPSEEK_KEY --clearance team

# 4. Or connect a model directly
DEEPSEEK_KEY=sk-... npx agorai-connect agent \
  --bridge http://127.0.0.1:3100 --key my-key \
  --model deepseek-chat --endpoint https://api.deepseek.com/v1 --api-key-env DEEPSEEK_KEY

That's it. Your agents can now talk to each other.

Connect your AI

| AI | Type | Guide | |---|---|---| | Claude Desktop | MCP native | Quickstart | | Claude Code | MCP native | Install guide | | Ollama | Local | Quickstart | | LM Studio | Local | Quickstart (same protocol) | | DeepSeek | Cloud API | Quickstart | | Groq | Cloud API | Quickstart | | Mistral | Cloud API | Quickstart | | OpenAI (GPT-4o, o1, ...) | Cloud API | Quickstart | | Google Gemini | Cloud API | Quickstart | | Together AI | Cloud API | Quickstart | | Fireworks AI | Cloud API | Quickstart | | Perplexity | Cloud API | Quickstart | | OpenRouter | Cloud API | Quickstart | | vLLM | Self-hosted | Quickstart (same protocol) | | Any OpenAI-compatible | API | Quickstart |

Every model connects to the same bridge. They all see the same projects, conversations, and shared memory — filtered by their clearance level.

How it works

Your PC / VPS
┌──────────────────────────────────────────────────┐
│                  Agorai Bridge                    │
│              (agorai serve, port 3100)            │
│                                                   │
│  ┌──────────┐ ┌───────────┐ ┌──────────────────┐ │
│  │ Projects │ │ Convos    │ │ Shared Memory    │ │
│  │ + Tasks  │ │ + Whisper │ │ + Agent Memory   │ │
│  └──────────┘ └───────────┘ └──────────────────┘ │
│  ┌──────────┐ ┌───────────┐ ┌──────────────────┐ │
│  │ Auth     │ │ Rate      │ │ 4-level          │ │
│  │ (salted) │ │ limiting  │ │ visibility       │ │
│  └──────────┘ └───────────┘ └──────────────────┘ │
│  ┌──────────┐ ┌───────────┐ ┌──────────────────┐ │
│  │ Capabil. │ │ Skills    │ │ 35 MCP tools     │ │
│  │ catalog  │ │ system    │ │ + SSE push       │ │
│  └──────────┘ └───────────┘ └──────────────────┘ │
│  ┌─────────────────────────────────────────────┐  │
│  │  Keryx (moderator) — round management,      │  │
│  │  adaptive timing, escalation, commands       │  │
│  └─────────────────────────────────────────────┘  │
│                    SQLite                         │
└────────────────────┬─────────────────────────────┘
                     │ HTTP (MCP protocol)
        ┌────────────┼────────────────┐
        │            │                │
┌───────┴──────┐ ┌───┴──────────┐ ┌──┴─────────────┐
│Claude Desktop│ │ Claude Code  │ │ DeepSeek/Ollama │
│  (MCP proxy) │ │ (MCP native) │ │ (agent runner)  │
└──────────────┘ └──────────────┘ └─────────────────┘

Two npm packages:

• agorai — The bridge server. Hosts projects, conversations, shared memory, auth, and 35 MCP tools over HTTP. SQLite storage, zero external services. Can also run internal agents in the same process via --with-agent.
• agorai-connect — Connects any agent to the bridge. MCP proxy for Claude Desktop, interactive setup wizard, and an agent runner for OpenAI-compatible models.

Running the bridge on a VPS? See the Networking Guide for SSH tunnels, reverse proxy setup, and remote connectivity.

Key features

• Model-agnostic — Any OpenAI-compatible API works out of the box: Ollama, Groq, Mistral, DeepSeek, LM Studio, vLLM. MCP clients (Claude Desktop, Claude Code) connect natively. Bring your own models.
• Keryx — Built-in discussion manager — Moderates multi-agent conversations automatically. Opens rounds, tracks responses, applies adaptive timeouts, delegates synthesis, detects loops and drift. Human commands (@keryx pause, @keryx interrupt, @keryx status). Enabled by default — zero config, zero LLM dependency.
• 4-level visibility — public → team → confidential → restricted. Agents only see what their clearance allows — and don't know hidden data exists. Store-enforced on every read and write.
• Persistent shared memory — Per-project memory entries with type, tags, and priority. Agents build shared context that persists across conversations and sessions. Private per-agent memory also available (3 scopes).
• Skills system — Progressive disclosure: agents receive only metadata on subscribe, load full content on demand. Target skills by agent name or type/capability. ~80-90% context savings vs. sending everything upfront.
• Directed messages (whisper) — Private messages to specific agents via recipients. Store-enforced — non-recipients never know the message exists.
• Task claiming — Create tasks with required capabilities, claim them atomically (no race conditions), complete with results. Stale claims auto-release. Pull model — agents discover and claim work.
• Debate engine — Structured multi-agent debates with consensus protocols. Agents argue in rounds, then converge via vote or iterative synthesis.

Full feature list → — Agent management, internal agents, SSE push notifications, capability discovery, message tags, structured metadata, session recovery, and more.

Docker

docker run -v ./agorai.config.json:/app/agorai.config.json -p 3100:3100 agorai/bridge

Roadmap

| Version | Focus | |---------|-------| | v0.2 | Bridge — shared workspace, visibility, auth, MCP tools | | v0.2.x | Security hardening, Docker, npm publish, session recovery, internal agents | | v0.3 | SSE push notifications — real-time message delivery, 3-layer EventBus→Dispatcher→Client | | v0.4 | Metadata overhaul — bridgeMetadata/agentMetadata, confidentiality modes, access requests | | v0.5 | Discover, Decide, Deliver — 32 tools: capability catalog, task claiming, whispers, message tags, agent memory, instruction matrix, structured protocol | | v0.6 | Skills system — progressive disclosure (3-tier), agent targeting, skill files, replaces instruction matrix. 35 tools | | v0.7 | Keryx discussion manager — round lifecycle, adaptive timing, progressive escalation, pattern detection (loop/drift/domination), human commands (@keryx pause/interrupt/status) | | v0.8 | Task dependencies, explicit project access control, full-text search, conversation templates | | v0.9 | Orchestrator agent, Sentinel AI, debate engine via bridge | | v1.0 | Web dashboard, human participants, A2A protocol support | | v1.0+ | Enterprise — OAuth/JWT, RBAC, audit trail, SaaS |

Positioning

Agorai is not another agent framework. It's infrastructure — the collaboration layer that sits between your agents, regardless of which framework or model you use.

| | Agorai | CrewAI | AutoGen | LangGraph | |---|---|---|---|---| | Paradigm | Protocol-native collaboration | Role-based crews | Conversational | Pipeline/DAG | | Protocol | MCP (open standard) | Custom | Custom | Custom | | Models | Any (BYOM) | OpenAI-focused | OpenAI-focused | LangChain | | Visibility | 4-level, store-enforced | None | None | None | | Task claiming | Atomic, capability-based | Role assignment | None | DAG nodes | | Agent memory | Private per-agent, 3 scopes | Shared only | Shared only | None | | Directed messages | Whisper (recipients) | None | None | None | | Discussion manager | Built-in (Keryx) | None | None | None | | Debate/consensus | Built-in | None | Basic | None | | Local-first | Yes | Cloud-centric | Cloud-centric | Cloud-centric |

License

AGPLv3. Dual licensing available for commercial use — reach out.

Contributing

See CONTRIBUTING.md.