The runtime

Ten packages. One capable agent.

Each piece is small, typed, and tested — and composes into the agentic-OS kernel. Nothing is a black box.

@anomalithic/providers

Any model

Anthropic, OpenAI, OpenRouter, Ollama, or any OpenAI-compatible endpoint.

@anomalithic/core

Signed thinking-impressions

Every thinking window mints a runtime-signed proof — the trust anchor for ads.

@anomalithic/mcp

Model Context Protocol

A first-class MCP client; bring any server's tools into the agent.

@anomalithic/skills

Claude + Codex skills

Loads Claude SKILL.md and Codex AGENTS.md into one skill system.

@anomalithic/hooks

Lifecycle hooks

Gate, mutate, or extend every step: SessionStart, Pre/PostToolUse, Stop, Thinking.

@anomalithic/orchestrator

Teams of agents

Durable task store, atomic checkout, budgets — run for hours or days.

@anomalithic/memory

Cross-session memory

A file-backed memory index plus recall, persistent across sessions.

@anomalithic/security

Guardrails

Secret redaction, permission policy, path sandbox, audit trail.

An agentic OS

A kernel for autonomous work.

The kernel wires providers, tools, skills, hooks, memory, and orchestration into one runtime you can run for minutes or for days — locally, on your keys.

Runs on
ClaudeGPTLlamaQwenDeepSeekOllamaOpenRouter
kernelcomposes everything below
providersany open or closed LLM
tools · mcplocal tools + MCP servers
skillsClaude SKILL.md + Codex AGENTS.md
hookslifecycle gates
memorycross-session recall
orchestratorteams of agents
securityredaction · permissions · sandbox
Runs anywhere

Ships as a tool, runs as a service.

anomalithic — session
$ anomalithic run "explain MCP in one line" --ads
✦ thinking…
💡 Your ad here while agents think — anomalithic.dev/ads
MCP is an open protocol that lets agents call external tools and
data sources over a uniform JSON-RPC interface.
[anthropic:claude-sonnet-4-6] 1 turn · 1 signed impression