Applications that run on amplifier-agent. Each one is a different answer to "what does a host look like", and each is a working reference for anyone writing their own integration.
Runs the opencode coding TUI on top of a local amplifier-agent. There is no fork: amplifier-opencode discovers which models your engine serves, writes an opencode config from that discovery, and launches stock opencode against it, re-syncing on every launch.
curl -fsSL https://raw.githubusercontent.com/microsoft/amplifier-app-opencode/main/install.sh | bash
amplifier-opencode launchIntegration shape: HTTP face. amplifier-app-opencode spawns amplifier-agent serve chat-completions and registers it as an opencode provider, so opencode itself only ever speaks the OpenAI protocol.
Sets up paperclip with the amplifier_local adapter so paperclip agents run on the engine. The paperclip server invokes amplifier-agent once per turn.
Integration shape: per-turn subprocess with one host config file per agent instance. This is the reference for the multi-tenant pattern described in the integration guide.
Runs NanoClaw, which routes chat channels into per-agent Docker containers, with amplifier-agent as the agent backend instead of its default. Selecting it writes NANOCLAW_DEFAULT_PROVIDER=amplifier-agent, which lets you choose between Anthropic, OpenAI, Azure OpenAI, and Ollama per agent group while keeping NanoClaw's sandbox, channels, and security model unchanged.
Integration shape: a Node host with the engine running inside each per-agent container.
Start at the integration guide. It opens with embedding the engine library, the primary surface, then covers the wrappers for hosts that cannot embed (TypeScript SDK, Python SDK, HTTP face, raw CLI contract), and ends with a checklist for each path. Each reaches the engine out-of-process for its own reason: opencode is an existing harness, nanoclaw is Node, and paperclip runs the engine in a container. A new host should reach for the library first.