Docs
In the agents (MCP + hooks)
The zero-UI path — run claude and codex as usual, they read the shared layer.
The simplest way to use uac is to not change how you work at all. After
uac init, running claude or codex in the project just works: they reach the
shared layer through MCP tools and session hooks.
What uac init wires up
| File | For | What it does |
|---|---|---|
.mcp.json | Claude Code | registers the uac MCP server (UAC_SOURCE=claude-code) |
.claude/settings.json | Claude Code | SessionStart + Stop hooks, tool permissions |
.claude/skills/ | Claude Code | generated native skill mirror |
.codex/config.toml | Codex | registers the MCP server (UAC_SOURCE=codex) |
.codex/hooks.json | Codex | SessionStart + Stop hooks |
CLAUDE.md | Claude Code | generated — @AGENTS.md + Claude extras |
~/.agents/registry.toml | both | makes the project linkable by name |
Attribution works because each agent spawns its own MCP server process, so
UAC_SOURCE is set per agent.
The two hooks
- SessionStart injects
project_context(AGENTS.md + skills index + recent memories) so the agent starts informed. - Stop nudges the agent to save durable facts via
memory_writebefore it finishes. Its own model does the writing — no API key.
Run the MCP server directly
Rarely needed, but available:
uac serve # stdio (what the agents use)
uac serve --http # streamable HTTP, for one long-lived serverHeadless / one-shot
uac run claude-code "explain what this repo does"
uac run codex "review the auth flow"
uac review "add retry logic to the client" # Claude writes → Codex reviews the diff