Agent Runtimes

matic does not call models directly. Instead, it delegates inference, tool use, and context materialization to pluggable Agent Runtimes such as Claude Code, Codex, Gemini CLI, and OpenClaw through a unified adapter interface.

Changing an agent's runtime changes how that agent thinks and acts, not who it is. Identity, memory, and accumulated experience remain stable across runtime changes, while the runtime layer controls execution behavior, environment preparation, streaming, and output normalization.

Adapter Interface

The single contract every runtime must implement, covering validation, environment preparation, tool provisioning, agent materialization, invocation, streaming, and output normalization. Read more.

Base Runtime

The abstract base class that all adapter implementations extend, providing shared retry logic, error normalization, cost tracking hooks, and stream buffering. Read more.

Claude Code

The default and most fully integrated runtime, covering hooks integration, CLAUDE.md injection, MCP surface exposure, worktree isolation, permission policy, and spawning lifecycle. Read more.

Codex

Adapter for OpenAI's Codex CLI, translating matic's context envelope and tool manifest into Codex-native form. Read more.

Gemini CLI

Adapter for Google's Gemini CLI, bridging matic orchestration to Gemini's execution model. Read more.

OpenClaw

Adapter for OpenCode / OpenClaw, integrating an open-source runtime option into the matic adapter framework. Read more.

Runtime Registry

The lookup layer that maps runtime slugs to adapter classes, making the invocation pipeline runtime-agnostic. Read more.