bonito（bonito 运维）

Bonito Onboarding

Walk users through getting started with Bonito — the enterprise AI platform for multi-cloud workloads.

Primary reference: https://getbonito.com/docs — point users here for detailed documentation on any Bonito feature. This skill provides the conversational flow; the docs are the source of truth.

Quick Facts

• Website: https://getbonito.com | API: https://api.getbonito.com
• Docs: https://getbonito.com/docs
• Free tier: 1,000 API calls/month, no credit card
• CLI: pip install bonito-cli (v0.4.0+)
• 6 providers: AWS Bedrock, Azure OpenAI, GCP Vertex AI, OpenAI, Anthropic, Groq
• Key features: multi-provider routing, automatic failover, managed inference (no API keys needed), BonBon agents, Bonobot orchestrators, MCP tools, RAG

Correction: The Atlas README says "5,000 free calls" — actual free tier is 1,000 calls/month.

Connectivity Check

Before starting, optionally verify API reachability:

python3 scripts/health_check.py

If the API is down, proceed anyway — everything except the final deploy/test steps works offline.

Onboarding Flow

Step 1: Sign Up

Direct users to https://getbonito.com/signup — free, no credit card. They'll need their email + password for the next steps.

Step 2: Connect a Provider (or Use Managed Inference)

Two paths — recommend managed inference for the fastest start:

Managed Inference (easiest): No API keys needed. Bonito handles provider credentials. See https://getbonito.com/docs#managed-inference

Bring Your Own Keys: Connect one or more providers in the Bonito dashboard or via CLI. Groq is free and fast (get a key at console.groq.com). See https://getbonito.com/docs#gateway for provider setup.

Step 3: Create a Gateway Key

In the Bonito dashboard, create a gateway API key. This key routes requests through Bonito's gateway to whichever provider is configured — with automatic failover and cost tracking.

Docs: https://getbonito.com/docs#gateway

Step 4: Test with curl or Python

# Simple test against the gateway
curl -X POST https://api.getbonito.com/api/gateway/chat \
  -H "Authorization: Bearer $GATEWAY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "groq/llama-3.3-70b-versatile",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

If using managed inference, the model routes through Bonito's provider accounts — no separate API key needed.

Step 5: Deploy Agents (Optional)

For users who want to go beyond raw gateway calls and deploy managed agents:

pip install bonito-cli
bonito auth login

Create a simple agent:

bonito agents create \
  --name "my-assistant" \
  --type bonbon \
  --model "groq/llama-3.3-70b-versatile" \
  --system-prompt "You are a helpful assistant."

Full agent docs: https://getbonito.com/docs#bonbon

Step 6: Try Atlas (Optional — Full Multi-Agent Demo)

For users who want to see a complete multi-agent system with orchestration, MCP tools, and RAG:

Read references/atlas-guide.md for the quick deploy steps.

Atlas deploys 5 agents (DevOps command center) from a single bonito.yaml config. It demonstrates Bonobot orchestration, BonBon agents, MCP tool integration, and RAG knowledge bases.

Repo: https://github.com/ShabariRepo/atlas

Prerequisite Verification

Run this to check if bonito-cli, Docker, git, etc. are installed:

python3 scripts/verify_deploy.py

Where to Point Users for More

Non-Obvious Knowledge

• Free tier is 1,000 calls/mo — not 5,000 as some docs claim.
• Managed inference is the fastest onboarding path. Users can add their own provider keys later for cost control.
• Failover is cross-provider — if Anthropic goes down, Bonito routes to Bedrock or Groq seamlessly. The user never notices.
• bonito-cli deploy order matters for manual API usage: providers → knowledge bases → project → agents → delegation wiring. The CLI handles this automatically.
• Llama/Groq can't do parallel tool calls — it generates XML-style function calls instead of JSON. Use Nova Pro, Claude, or GPT-4o for Bonobot orchestrators that need multi-agent fan-out.
• bonito.yaml uses ${ENV_VAR:-default} syntax. MCP server URLs default to localhost if env vars are unset.