How can we help?

AgentNode is a verified registry and runtime for AI agent capabilities. Agents discover, install, and run verified packages — tool packs, MCP servers, and agents — at runtime, governed by trust tiers, policy checks, and lockfiles.

Four audiences: people who want to give an agent new capabilities, developers who publish packages, teams who must govern what agents run, and AI agents themselves, which can read a machine-readable setup guide.

No. AgentNode is a registry plus a runtime that works with LangChain, CrewAI, MCP, and plain Python. You keep your framework; AgentNode supplies and governs the capabilities.

Install the SDK with pip install agentnode-sdk, run agentnode setup, then search, install, and run a package. No account is needed to search, install, or run.

Tool packs, Skills, MCP servers, and full Agents — all in the portable ANP package format, discoverable by capability.

A tool pack provides callable tool functions; a skill provides prompt templates and assets; an MCP server is an external Model Context Protocol tool process; an agent orchestrates an LLM and tools toward a goal.

Scaffold with agentnode init, validate it, run agentnode verify-local, then agentnode publish. Publishing requires an AgentNode account and an API key; browsing, installing, and running do not.

No account or key is required to search, install, or run packages. You need a provider key only if you use a hosted LLM (such as OpenAI or Anthropic); an AgentNode API key is required only to publish.

Yes. Ollama runs models locally with no hosted API key — once Ollama is installed, running, and a model is pulled. AgentNode never installs or starts it for you.

In your OS keychain, or a local file with owner-only (0600) permissions as a fallback. An environment variable always overrides the stored value. Keys are not uploaded to AgentNode, though a hosted provider receives your prompts and responses when you use it.

Built-in: OpenAI, Anthropic, OpenRouter, DeepSeek, Mistral, Qwen, Gemini, plus local Ollama and custom OpenAI-compatible endpoints. The runtime is tested across 246 models from 40 providers (222 pass all scenarios).

Untrusted community packages run in a hardened container or not at all (fail-closed), when a container runtime and the pinned image are available. Trusted and curated packages run host-side with policy checks and subprocess isolation — not OS-level sandboxing.

An agent's own code runs host-side, so trusted and curated agents run on the host. Community agents are refused by default — they run only if you opt into the agent sandbox, which is off by default.

They are trust levels: unverified (metadata validated only), verified (security-scanned, publisher confirmed), trusted (proven reliable over time), and curated (reviewed by AgentNode). Package versions are also scored into verification tiers.

When runtime isolation is required, it is sandboxed-or-fail-closed: it runs in a container, or — if no sandbox is available — it is blocked, never silently run on the host.

Guard is a pre-execution policy gateway. Every install and run is checked against trust level, permissions, and environment, and tool actions are classified allow / prompt / deny with rate limits. It is fail-closed.

AgentNode does not currently claim SOC 2, HIPAA, ISO, or similar compliance certifications. It is local-first, with trust tiers, publisher signatures, lockfile integrity, and sandbox-or-fail-closed isolation for untrusted code.

Through machine-readable files: /llms.txt and a deeper /llms-full.txt guide that explains setup, security boundaries, and the install commands for AI agents.

Start with Getting Started and the Quick Start, then the security model and the CLI reference.