Adding a New Host

This page consolidates the checklists from the Adapter Playbook and the Adapters README.

Design First

From the Adapter Playbook:

• Can the host rewrite requests before model execution?
• Can the host rewrite responses after model execution?
• Does the host expose streaming chunks, final responses, or both?
• Does the host expose transcript history directly, or must it be reconstructed?
• Is integration file-based, hook-based, plugin-based, or API-based?

Required Surfaces

From the adapters/README.md checklist:

1. Install surface — where the host extension or plugin lives, how config is written, how to enable/disable/remove.
2. Session bridge — how session IDs, turn IDs, and workspace roots are resolved.
3. Transcript bridge — how raw host messages are decoded into shared envelopes and rewritten envelopes encoded back.
4. Request / response lifecycle — before-call rewriting, after-call reduction, tool-result persistence, streaming/non-streaming handling.
5. State roots — state dir, archive dir, debug trace dir.
6. Product surface — status, report, doctor, visual, mode switching.
7. Verification — adapter unit tests, one smoke install path, one end-to-end session check.

Recommended Development Order

From the adapters/README.md:

1. Start with the shared host envelope bridge in packages/host-adapter/.
2. Wire one minimal request path through the host.
3. Make status and doctor work first.
4. Add reduction and persistence hooks.
5. Add visual and richer command surfaces last.

Reference Implementations

• OpenClaw Adapter
• Codex Adapter
• Claude Code Adapter

Related Pages

• Adapter Architecture
• Adapter Testing
• Configuration Integration
• Hook and Proxy Integration
• Adapter Playbook
• HOSTS.md