Bonito MCP Server

A Model Context Protocol (MCP) server that wraps the Bonito AI Platform REST API. Use it from Claude Desktop, Cursor, or any MCP-compatible client to manage your multi-provider AI infrastructure.

Features

• Provider Management – Connect and verify AWS Bedrock, Azure, Google Vertex, OpenAI, Anthropic, Groq, OpenRouter and DeepInfra
• Model Management – List (by provider, search, chat-only), sync, and activate models
• Unified Gateway – Chat completions through Bonito's OpenAI-compatible gateway; gateway keys limited to providers or models (create, update, revoke)
• Projects – List and create projects (budget, retention)
• Agents – Create, update, run (with sessions) and disable agents; read a run that was queued at the agent's rate limit
• Knowledge Bases – Create KBs, upload text documents, list their status, search them, link and unlink them to agents
• Cost & Observability – Gateway usage, request logs, and cloud billing per provider
• Data Residency – Keep model calls in one geography (CA, US, EU) and see which providers can serve (get_data_residency, set_data_residency)
• Stacks (bonito.yaml) – Plan a whole stack (providers, routing, keys, knowledge bases, agents, MCP servers, integrations) and apply it after review (plan_stack, apply_stack)
• Agents in your clouds – List agents running in your own AWS and Google accounts (AgentCore, Bedrock Agents, Agent Registry, Agent Engine) with findings, let Bonito agents call them, and manage the relay that lets cloud agents call each other (list_cloud_agents, connect_cloud_agent, agent_relay, Pro+)

Total: 38 tools. 1.5.0 adds plan_stack and apply_stack: give a bonito.yaml (as text or a path on your machine) and Claude shows what would be created, changed, refreshed or left alone, then applies exactly that plan once you agree. See Stacks below. 1.4.2 fixes the write tools: create_agent sent fields the API rejects (422), create_knowledge_base left out source_type, upload_to_kb posted JSON to a file-upload endpoint, link_kb_to_agent called a route that does not exist, and several tools sent parameters the API ignores (create_project budget, list_models filters, gateway_usage and get_costs periods). Errors now carry the API's own message, and empty (204) replies no longer crash. New: update_agent, delete_agent, unlink_kb_from_agent, search_knowledge_base, list_kb_documents, revoke_gateway_key. 1.4.1 fixed every tool failing with "ctx: Field required".

Use a bp- personal access token as BONITO_API_KEY: it works on both the management API and the gateway (chat_completion, ask_decision). A sign-in token works for management calls only.

Still missing (next pass): access-token CRUD, image and video generation, and agent-to-agent connections outside a stack.

Quick Start

Install from source

git clone https://github.com/ShabariRepo/bonito.git
cd bonito/mcp-server
pip install -e .

Set your API key

export BONITO_API_KEY="your-bonito-api-key"

Run (stdio – for Claude Desktop)

bonito-mcp

Run (SSE – for network clients)

bonito-mcp --transport sse --port 8080

Claude Desktop Configuration

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "bonito": {
      "command": "bonito-mcp",
      "env": {
        "BONITO_API_KEY": "your-bonito-api-key"
      }
    }
  }
}

Docker

docker build -t bonito-mcp .
docker run -e BONITO_API_KEY="your-key" -p 8080:8080 bonito-mcp

Available Tools

Stacks (bonito.yaml)

plan_stack takes a bonito.yaml as yaml_text or as a path on the machine running the MCP server. Prompt files, knowledge base directory sources and service_account_file entries it points at are read there and sent with the file (20 MB in all). Only files inside the yaml's folder are read; absolute, ~ and ../ paths that leave it are reported in not_found_locally and never sent. The reply is the plan: each resource with its action (create, update, refresh, unchanged), the fields that change, and any errors, warnings or unresolved variables, plus a plan_hash.

apply_stack takes that plan_hash with the same file. If anything changed on the platform since the plan, nothing is applied and the reply has code: "plan_changed" and the fresh plan to review. Apply creates and updates; it never deletes. A new gateway key's or agent token's secret appears once, in new_keys (agent tokens carry kind: "project token").

Keep secrets out of the file as ${VARS}. A variable is sent only if the env argument gives it, or if its name is listed in BONITO_STACK_ENV (comma-separated) and its value is set in the MCP server's environment (the env block of your client config). BONITO_* variables are never sent. Names the file uses that were not sent come back in env_not_sent. The file decides nothing about what is read from your machine, so a yaml Claude was shown cannot pull other secrets.

Environment Variables

License

MIT