Nanobot – Turn MCP servers into full AI agents

TL;DR

TL;DR: Nanobot is an open-source, standalone MCP host that turns any MCP server into a dedicated AI agent, configurable via a single nanobot.yaml file and runnable in minutes.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: nanobot.ai ← visited and verified
• Source repository: github.com/obot-platform/nanobot ← read README end-to-end
• License: Apache 2.0 ← verified from LICENSE file on main branch
• HN launch thread: news.ycombinator.com/item?id=45294107
• Source last checked: 2026-06-24 (main branch, commit verified via raw README fetch)

What Is Nanobot?

Every AI assistant you use — Claude, ChatGPT, Cursor, Goose — ships with an embedded MCP host. It connects your LLM to external tools and data sources. But that host is tied to that product. You cannot easily take Claude’s MCP host and run it independently, plug it into your own app, or deploy it as a standalone service.

Nanobot solves this. It is a standalone, open-source MCP host you deploy yourself. Point it at any MCP server, pick a model, write a YAML config, and you have a dedicated agent running at http://localhost:8080.

From the README:

“Nanobot enables building agents with MCP and MCP-UI by providing a flexible MCP host. While existing applications like VSCode, Claude, Cursor, ChatGPT, and Goose all include an MCP host, Nanobot is designed to be a standalone, open-source MCP host that can be easily deployed or integrated into your applications.”

Think of it as the missing runtime for MCP — the thing that runs between your LLM and your tools, extracted into a portable, self-hosted binary.

Setup Workflow

Step 1: Install

The fastest route is Homebrew:

brew install obot-platform/tap/nanobot

This installs the nanobot CLI. Alternatively, build from source:

git clone https://github.com/obot-platform/nanobot.git
cd nanobot
make

Step 2: Configure your agent

Create a nanobot.yaml file. Here is a minimal example that connects a Blackjack Dealer agent to a remote MCP server:

agents:
  dealer:
    name: Blackjack Dealer
    model: gpt-4.1
    mcpServers:
      - blackjackmcp

mcpServers:
  blackjackmcp:
    url: https://blackjack.nanobot.ai/mcp

Nanobot supports two model providers out of the box:

• OpenAI — set OPENAI_API_KEY, use model: gpt-4.1
• Anthropic — set ANTHROPIC_API_KEY, use model: anthropic/claude-3-7-sonnet-latest

It auto-selects the provider from the model string. For other providers (Azure, Bedrock, Ollama), add an llmProviders block with a dialect and base URL.

Step 3: Run

nanobot run ./nanobot.yaml

The web UI starts at http://localhost:8080.

Directory-Based Configuration

For multi-agent setups, use a directory layout instead:

my-config/
├── nanobot.yaml        # Shared: mcpServers, llmProviders, env
└── agents/
    ├── main.md         # Primary agent (entrypoint)
    └── helper.md       # Additional agents

The nanobot.yaml holds shared infrastructure; each .md file in agents/ defines an individual agent. Markdown agents override same-named entries in nanobot.yaml, so you can override shared config per-agent.

Agent markdown files use YAML front-matter for config and the body as system instructions:

---
name: Shopping Assistant
model: anthropic/claude-3-7-sonnet-latest
mcpServers:
  - store
temperature: 0.7
---

You are a helpful shopping assistant.

Help users find products and answer their questions.

Deeper Analysis

Why this matters for the MCP ecosystem. MCP has a discovery problem: great servers exist (Filesystem, GitHub, Hugging Face, Shopify) but most people only access them through one specific tool. Nanobot decouples the host from the client, making any MCP server a reusable building block for your own agents.

Multi-provider flexibility. The llmProviders block with dialect selection is notable. The OpenResponses dialect covers any OpenAI-compatible endpoint, which includes Ollama, LM Studio, Jan.ai, and cloud providers that expose OpenAI-compatible APIs. This makes Nanobot viable as a local-first stack:

llmProviders:
  ollama:
    dialect: OpenResponses
    baseURL: http://localhost:11434/v1

mcpServers:
  store:
    url: https://example.com/mcp

Set OPENAI_API_KEY or ANTHROPIC_API_KEY for cloud models with zero additional config — the provider activation is automatic.

UI as a feature, not the product. The built-in web UI at port 8080 is convenient for demos and local use. But the real value is the nanobot run binary as infrastructure you embed or deploy. The roadmap mentions Slack, SMS, email, and web embedding — positions Nanobot as a middleware layer rather than a chatbot.

Practical Evaluation Checklist

• Installation: Homebrew one-liner, no Docker required for local use
• Config format: YAML, human-readable, no code needed
• Provider support: OpenAI, Anthropic, Azure, Bedrock, Ollama (via dialect config)
• MCP server support: Any server that exposes an HTTP endpoint with the MCP protocol
• Multi-agent: Yes, via directory-based config with markdown agent files
• Local-first: Fully offline capable with Ollama
• Development status: Alpha — the README warns of active breaking changes
• License: Apache 2.0

Security Notes

• API keys are injected via environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY), not hardcoded in the YAML
• MCP server URLs and headers (including auth tokens) are defined in the config file — treat your nanobot.yaml as a secrets file
• The web UI (port 8080) has no built-in auth — run behind a reverse proxy if exposing externally

FAQ

Q: How is this different from using Goose or Claude’s MCP host? A: Goose and Claude ship with an embedded MCP host tied to their UX. Nanobot is the host extracted as a standalone binary. You control the config, the model, and the deployment target. It is infrastructure, not a product.

Q: Can I use local models (Ollama)? A: Yes. Add an llmProviders entry with dialect: OpenResponses and baseURL: http://localhost:11434/v1, then set your model to any Ollama model name.

Q: Does it work with MCP-UI (the newer protocol)? A: Partial support — the roadmap lists MCP-UI as a work-in-progress feature.

Q: Is this production-ready? A: No — the project is in alpha and the README explicitly warns of breaking changes. It is not suitable for production deployments without pinning to a specific commit.

Conclusion

Nanobot fills a gap in the MCP ecosystem: there are dozens of MCP servers but no standard, portable runtime to wire them to arbitrary models. If you want to prototype a dedicated agent backed by, say, a Shopify or Hugging Face MCP server, Nanobot gets you there in five minutes with a YAML file.

The alpha status is the main caveat — do not deploy it as a permanent infrastructure component without a pinning strategy. Watch the repo for 1.0 signs. For local development and experimentation, it is already genuinely useful.

Source and Accuracy Notes

• Project page: nanobot.ai
• Source repository: github.com/obot-platform/nanobot
• License: Apache 2.0 (verified via LICENSE file on main branch)
• HN launch thread: news.ycombinator.com/item?id=45294107