adapters

adapters/

Drop-in glue for connecting vouch to specific LLM hosts.

Each subdirectory is a small, self-contained example of wiring vouch serve into one host. They are templates, not packages — copy the file you need into your project and edit it.

Adapter	Host	Files
claude-code/	Anthropic's Claude Code CLI	.mcp.json snippet, CLAUDE.md excerpt
cursor/	Cursor IDE	mcp.json snippet
codex/	OpenAI's Codex CLI	tiered install: .codex/config.toml merge, AGENTS.md excerpt, skills, capture hook
continue/	Continue.dev	config.json snippet
openclaw/	OpenClaw plugin host	.openclaw/plugins.json, AGENTS.md excerpt
generic-mcp/	Any MCP-speaking host	annotated reference
jsonl-shell/	bash scripts via the JSONL transport	example pipeline

What an adapter is responsible for

Everything host-specific. That includes:

1. How to launch vouch serve with the right transport and env.
2. What identity to give the agent. Set VOUCH_AGENT distinctly per host so the audit log can attribute writes.
3. Where to put the host's configuration file. Each host has its own preferred location.
4. Tool-name mapping. MCP hosts see methods as kb_search, kb_propose_claim etc. — see spec/transports.md.

What an adapter is not responsible for

• The KB's contents — those live in .vouch/ regardless of host.
• Approval workflow — that's vouch approve from your shell. No adapter routes approval through the agent.
• Method semantics — the kb.* surface behaves identically in every host.

Adding an adapter

Open a PR. New adapter directories should contain:

• A short README.md explaining where the snippet goes and what changes to make.
• The minimal config snippet itself (one file).
• Optional: an agent-prompt.md with phrasing tips for that host's conventions (e.g. "Claude Code reads CLAUDE.md; here's a paragraph to drop in").

No code. Adapters are configuration, not packages — anything that needs Python belongs in the core.