tooti 🦜

A decentralized AI inference protocol with an OpenAI-compatible gateway.

tooti connects model providers and application developers through a shared protocol layer:

• operators run nodes and advertise available models
• the network discovers healthy providers
• the gateway routes requests to the best available node
• clients use a familiar OpenAI-style API

Why tooti

AI apps need reliable inference, but today teams often face:

• single-provider dependency
• regional outages and latency spikes
• limited control over routing and cost
• difficult migration paths between model backends

tooti exists to make inference portable, resilient, and open:

• portable for developers (standard API, minimal lock-in)
• resilient for production (multi-node discovery and failover)
• open for operators (bring your own backend: Ollama, vLLM, llama.cpp, and more)

Architecture

flowchart LR
    A[AI App / Client] -->|OpenAI-compatible HTTP| B[Tooti Gateway]
    B --> C[Router]
    C --> D[(Node Registry)]
    D --> E[Node Agent A]
    D --> F[Node Agent B]
    D --> G[Node Agent C]
    E --> H[Ollama / vLLM / llama.cpp]
    F --> I[Ollama / vLLM / llama.cpp]
    G --> J[Ollama / vLLM / llama.cpp]
    E <-. libp2p heartbeat + capabilities .-> D
    F <-. libp2p heartbeat + capabilities .-> D
    G <-. libp2p heartbeat + capabilities .-> D

Loading

What each part does

• Node Agent: advertises model capabilities, health, and availability over libp2p
• Registry + Router: tracks live nodes and selects where each request should go
• Gateway: exposes /v1/models and /v1/chat/completions for client apps

Network bootstrap

Configure network.bootstrap_peers (e.g. /dnsaddr/discover.tooti.network to expand peer ids from TXT, or a full /ip4/.../p2p/...).

Quick start

1) Build

go build -o tooti ./cmd/tooti

2) Validate and run

# Validate config first
./tooti config-check -file ./node.yaml

# Start node
./tooti node start -file ./node.yaml

# Start gateway (same host or another host)
./tooti gateway start -file ./node.yaml

3) Test the API

# List models
curl -s http://127.0.0.1:8080/v1/models

# Stream chat completion
curl -N http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"llama3.2:latest","stream":true,"messages":[{"role":"user","content":"say hi"}]}'

OpenClaw integration

• Guide: docs/openclaw.md
• Example config: docs/openclaw.json.example

Contributing

See CONTRIBUTING.md.

License

MIT