Nanobot
by Obot AI (formerly Acorn Labs)
Open-source MCP host that turns MCP servers into stateful agents
Nanobot is an open-source MCP host from Obot AI that turns a Model Context Protocol server into a complete agent. It supplies the runtime, stateful conversation threads, model routing and MCP-UI rendering around your existing MCP tools, all declared in a single YAML file and shipped as one Go binary.
Nanobot is an open-source MCP host from Obot AI — the company formerly known as Acorn Labs, built by the team behind Rancher and k3s — that turns a Model Context Protocol server into a complete, stateful agent. Most MCP servers only expose functions; Nanobot wraps them with everything else an agent needs, bundling an MCP client, an orchestrator, an LLM connection, a database and resource and user management into a single Go binary. Agents are declared either in one nanobot.yaml file or as markdown files in an agents/ directory, specifying the model, system prompt, starter messages, connected MCP servers and display metadata such as icons and descriptions; nanobot run ./nanobot.yaml then serves the agent over HTTP on localhost:8080. Because Nanobot itself speaks MCP, an entire agent can be published as a single MCP endpoint for agent-to-agent calls, collapsing what would otherwise be hundreds of individual tool definitions into one interface and cutting the context overhead that comes with them. It supports OpenAI, Anthropic, Azure, Amazon Bedrock and local Ollama models as backends, and includes MCP-UI, which lets a server return renderable interface resources so the human manipulates real UI while the model handles the logic — the team's demo was a blackjack dealer agent that keeps game state across turns. Announced on 18 September 2025, Apache-2.0 licensed, written in Go, with 1,334 GitHub stars and 204 forks. The README carries an explicit warning that the project is under heavy development and moving away from its original design and intent, so breaking changes should be expected.
The platform engineer who already runs MCP servers internally and wants to ship them as usable agents without adopting a proprietary client or writing another agent framework from scratch.
An existing MCP server becomes a working, stateful agent with a UI from a single YAML file and one Go binary — no Kubernetes, no SDK, no vendor account.
At a Glance
- Category
- AI Agents & Orchestration
- Pricing
- Free
- Target Market
- CTOs, Enterprise Developers, Platform Engineers, AI Engineers
- Deployment
- Open-source, Self-hosted
- Founded
- 2025
Key Features
- ✓Complete MCP host runtime
Bundles the MCP client, orchestrator, LLM connection, database and user management so an MCP server becomes a working agent.
- ✓YAML or markdown agent definitions
A single nanobot.yaml, or one markdown file per agent in an agents/ directory, defines the entire agent declaratively.
- ✓Stateful conversation threads
Conversation and tool context persist across turns, so multi-step workflows do not lose state between individual calls.
- ✓MCP-UI rendering
MCP servers can return interface resources that Nanobot renders, so users click real UI instead of trusting model-reported state.
- ✓Agent-to-agent over MCP
A whole agent is published as one MCP endpoint, replacing hundreds of tool definitions with a single standardised interface.
- ✓Multi-provider model support
Works with OpenAI, Anthropic, Azure, Amazon Bedrock and local Ollama models without changing agent code.
- ✓Single Go binary
Installs from the obot-platform Homebrew tap and runs with one command, requiring no Kubernetes or cloud account.
Capabilities
Use Cases
- •Turning an existing MCP server into a product
Wrap tools you already ship with reasoning, a system prompt and a UI to hand users a working agent.
- •Interactive tool UIs inside chat
MCP-UI lets a server render forms or boards the user manipulates directly, avoiding hallucinated state changes.
- •Composing multi-agent systems
Publishing each agent as an MCP endpoint cuts the context cost of exposing every underlying tool to every caller.
- •Local prototyping of agent behaviour
Run a complete agent on localhost against Ollama before committing to a hosted model or a managed platform.
- •Standardising internal agent definitions
Teams keep agent prompts and tool wiring in version-controlled YAML instead of scattered across application code.
Ideal For
Best For
- ✓Teams that already run MCP servers and want to ship them as usable agents
- ✓Platform engineers standardising internal agent definitions in version control
- ✓Developers who want a local, self-hosted MCP host rather than a proprietary client
- ✓Builders experimenting with MCP-UI and agent-to-agent composition over MCP
- ✓Go and Kubernetes shops already familiar with the Rancher and Acorn lineage
Not Ideal For
- ✗Production systems needing API stability — the README states the project is under heavy development and moving away from its original design and intent
- ✗Non-web channels: independent write-ups note the implementation is focused on web interfaces, with email, SMS and voice still aspirational rather than shipped
- ✗Buyers wanting a supported commercial product; nanobot.ai now redirects to obot.ai, whose homepage markets the MCP Gateway and does not mention Nanobot at all
- ✗Teams needing rich multi-agent orchestration today, since the project's own limitations list multi-agent capability as needing enhancement and MCP-UI support as partial
Integrations
Deployment
Market Analysis
Pros
- ✓Fills a real gap: MCP standardised tools but not the host, and Nanobot is a credible open-source implementation of one
- ✓Apache-2.0, a single Go binary and a Homebrew install make it trivial to try and trivial to remove
- ✓MCP-UI is genuinely differentiated — the human drives real UI while the model handles logic, which sidesteps a whole class of hallucination
- ✓Credible provenance in Darren Shepherd and the Rancher and k3s team, now operating as Obot AI
Cons
- ✗Self-declared alpha, with the README explicitly warning of breaking changes and architectural churn
- ✗MCP-UI support is only partial and multi-agent capability is listed by the project as needing enhancement
- ✗42 open issues and a small community, with no G2, Capterra or TrustRadius presence whatsoever
- ✗Vendor attention appears to have shifted: nanobot.ai now redirects to obot.ai, whose homepage markets the MCP Gateway and never mentions Nanobot
- ✗Its Show HN drew only 24 points and no comments, so there is very little public practitioner experience to learn from
Pricing
Open source (Apache-2.0)
$0
- ✓Full MCP host runtime
- ✓YAML and markdown agent config
- ✓MCP-UI support
- ✓Multi-provider LLM routing
- ✓Homebrew install
Free and Apache-2.0 licensed with no paid Nanobot tier at all; you pay only for the model tokens your agents consume, and Obot AI monetises separately through its hosted MCP Gateway and enterprise editions.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 4 on domains other than github.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Sema4.ai
Enterprise AI agents for knowledge work, running natively in your Snowflake or cloud account
Sapiom
Agent infrastructure that routes, runs and meters AI agents in production
Itential FlowAI
Governed AI agents for network and infrastructure operations, with deterministic execution and full audit trails
OpenAI Presence
Deploy production-grade AI voice and chat agents with enterprise policies, guardrails and evals