AgentLayer▸docs
Run it

Hosts

Your agent runs inside an agent CLI you already use. The home is the same on every host; what differs is how the host is wired to it.

A host is the agent CLI that opens the home, reads the manual, runs the skills, and calls the MCP server. Two are supported today: Claude Code and Codex. Everything the host does not own lives in the home, which is why a home works from either, and why a session ended in one is in the other's context next time.

What a host provides

PieceOwnerClaude CodeCodex
Operating manualHomeAGENTS.md through the .claude/CLAUDE.md bridgeAGENTS.md, read natively
SkillsPluginBundledBundled
MCP serverPluginRegistered by the plugin manifestRegistered per home in .codex/config.toml
HooksPluginBundledGenerated per home into .codex/hooks.json
Permissions and envHome.claude/settings.json, settings.local.jsonThe same files, projected into a profile, a rules file, and a shell env policy
Session capturePluginClaude transcript formatCodex rollout format

The home's settings live under .claude/ and are the source of truth for both hosts: Codex's wiring is generated from them, so a permission or a variable is set once. See Configuration.

Speaking to either host

Claude CodeCodex
Launch from the homeclaudecodex
Run a skill/agentlayer:sync$sync
InstallQuickstartCodex
Update the plugin/plugin update agentlayer@agentlayercodex plugin remove then codex plugin add (Updating)
Reconcile the home after an update/agentlayer:upgrade$upgrade
See what loaded/contextThe SessionStart payload names the host on its first line

These docs write skills by their bare name (sync, or briefing morning for a skill and one of its playbooks) and use host tabs wherever the exact command differs. Pick your host once in any tab strip; the choice sticks across pages.

The host does the thinking

The MCP server never calls a model. It does I/O and returns prompts; your own session synthesises, and the server confirms the write. The compile loop, briefings, and audits all run inside your turn, so they draw on whatever plan your host bills, never on a metered API key.

The consequence is that your agent is reactive: you start every session, and nothing runs at 7am on its own. A scheduler can run a host with no one at the keyboard, but those calls bill the pay-per-token API, not your plan. See Before the plugin for why. On Claude Code you can check it against your usage meters.

Adding a host

A new host needs four things, and the plugin isolates each behind a small surface:

  1. Read the manual. Any CLI that reads AGENTS.md gets the operating rules; the identity stack and indexes ride in on the SessionStart hook.
  2. Run the four hooks. session-start, session-capture at end and before compaction, and guard before shell commands. All run the plugin's CLI (bun "<plugin root>/src/cli.ts") with a --hook-protocol=<host> flag.
  3. Register the MCP server with the home pinned (AGENT_HOME in its env).
  4. A capture adapter. One format module that turns the host's transcript into user and assistant turns.

Codex was the first host added this way; its page is the worked example.

On this page