@agenticmail/codex

@agenticmail/codex

OpenAI Codex CLI integration for AgenticMail — surfaces every AgenticMail agent as a native Codex subagent and wires the dispatcher daemon to the Codex SDK so agents wake on their own mail and tasks.

This is the Codex CLI sibling of @agenticmail/claudecode. It uses the same dispatcher core (per-agent serialization, wake-coalesce, wake-budget, restart recovery, capabilities preamble) — only the host bindings differ.

What it does

Five things, one binary each:

• MCP server registration. Writes [mcp_servers.agenticmail] into ~/.codex/config.toml so the AgenticMail toolbelt (~60 tools under mcp__agenticmail__*) is available inside every Codex session.
• Multi-agent feature flag. Sets features.multi_agent_v2.enabled = true so Codex exposes the spawn_agent tool to the model.
• Subagent registration. Writes one TOML file per AgenticMail account into ~/.codex/agents/ so each agent is callable as spawn_agent({ agent_type: "agenticmail-<name>", message: "..." }).
• Lifecycle hooks. Registers the mail-hook on SessionStart / UserPromptSubmit / Stop in ~/.codex/hooks.json — injects the capabilities preamble on session start (including after auto-compact, where session_id survives but model context is wiped), surfaces fresh inbox mail before user prompts, and forces continuation on stop when there's bridge mail to act on.
• Dispatcher daemon. Long-running PM2-managed process that watches every AgenticMail account's IMAP IDLE stream and spawns a one-shot Codex turn (via @openai/codex-sdk) for each new mail or task event.

Install

npm install -g @agenticmail/codex
agenticmail-codex install

The install command writes config + agent files + hooks idempotently. Re-runs are no-ops when nothing has changed.

agenticmail-codex status         # is everything wired up?
agenticmail-codex uninstall      # remove everything we wrote
agenticmail-codex --help

You can also drive it programmatically:

import { install, uninstall, status } from '@agenticmail/codex';
const result = await install();

Architecture

Same shape as the Claude Code integration — only the host bindings differ:

	Claude Code	Codex
Global config	~/.claude.json (JSON)	~/.codex/config.toml (TOML)
Hooks file	nested in ~/.claude/settings.json	separate ~/.codex/hooks.json
Agent file format	markdown + YAML frontmatter	TOML with developer_instructions heredoc
Spawn tool	Agent({ subagent_type, prompt })	spawn_agent({ agent_type, message })
Feature gate	none	features.multi_agent_v2.enabled = true
SDK	@anthropic-ai/claude-agent-sdk	@openai/codex-sdk

The dispatcher's event loop reads Claude-shaped frames (assistant text, tool_use, tool_result, result). For Codex, a small adapter in dispatcher.ts:defaultQuery translates Codex's ThreadEvent stream into Claude-shaped frames so the rest of the dispatcher (channels, wakes, coalesce, budget, catch-up scan, dispatcher-state persistence) stays unchanged.

Codex's ThreadEvent → Claude frame mapping:

Codex event	Claude frame
item.completed(agent_message)	{ type: 'assistant', message: { content: [{ type: 'text', text }] } }
item.completed(mcp_tool_call)	{ type: 'assistant', ..., tool_use: mcp__<server>__<tool> } + { type: 'user', ..., tool_result }
item.completed(command_execution)	tool_use Bash + tool_result with aggregated_output
item.completed(file_change)	tool_use Edit with changes
item.completed(web_search)	tool_use WebSearch with query
turn.completed	{ type: 'result', result: <final_text>, usage }
turn.failed / error	{ type: 'result', result: '[error] ...', usage: {} }
reasoning / todo_list	dropped (no Claude equivalent surfaced today)

This keeps the read_email(uid) end-of-turn dedup logic (added in 0.9.5) working without modification — the tool_use breadcrumb format matches what the existing regex expects.

Environment overrides

Env var	What it overrides
AGENTICMAIL_API_URL	AgenticMail master API URL (default http://127.0.0.1:3829)
AGENTICMAIL_MASTER_KEY	Master key (otherwise read from ~/.agenticmail/config.json)
CODEX_HOME	Codex's home dir (default ~/.codex)
CODEX_AGENTS_DIR	Codex agents dir (default <CODEX_HOME>/agents)
AGENTICMAIL_DISPATCHER_MAX	Worker concurrency cap (default 50)
AGENTICMAIL_DISPATCHER_SYNC	Account sync interval ms (default 30000)

Status reporting

agenticmail-codex status

Prints a one-line state plus details:

• MCP server registered: is [mcp_servers.agenticmail] in ~/.codex/config.toml?
• multi_agent_v2 enabled: is spawn_agent exposed to the model?
• Bridge agent in AgenticMail: does the codex account exist on the master API?
• Subagent files: how many agenticmail-*.toml files are in ~/.codex/agents/?
• Dispatcher daemon: running? restart count? uptime?

status --json emits the same data as machine-readable JSON for CI / scripting.

Idempotency + ownership

Every file we write carries a # managed-by: @agenticmail/codex marker. The installer:

• never overwrites a user-authored agent file with the same name
• never touches keys in config.toml other than [mcp_servers.<our-name>] and features.multi_agent_v2.enabled
• preserves other users' hook entries on the same events
• prunes our agent files when their target AgenticMail account no longer exists

Uninstall reverses everything we wrote, leaving user state untouched. By default it keeps the bridge agent in AgenticMail (an account with an inbox, contacts, etc. shouldn't disappear silently); pass --purge-bridge to delete it.

External inbox exposure — what setup-email actually does to your dispatcher

Once the operator runs agenticmail setup-email, every Codex subagent on this machine becomes reachable from the public internet via Gmail / Outlook plus-addressing. Worth surfacing before the operator connects a relay:

• Plus-addresses are publicly guessable. Anyone can hit [e-mail], [e-mail], … and the matching subagent's inbox receives the mail. The +sub part is not a secret.
• External mail goes through the same handleEvent path as internal @localhost mail. Dedup, thread-cache, and wake-budget checks all run; if they pass, the Codex dispatcher spawns a fresh worker turn via @openai/codex-sdk to process the message. Source doesn't matter to the wake path.
• The bridge takes a different path on purpose. Mail to [e-mail] routes to handleBridgeMail, which uses codex.resumeThread(id).runStreamed(prompt) to wake the operator's last thread headlessly rather than spawning a new worker — so external mail to the bridge can wake your interactive CLI, not just background turns. If resume fails (thread expired, no host CLI running), it falls through to the bridge-escalation email at setup_operator_email.
• Spam wakes Codex turns. A scraper that finds a plus-address can drive billable Codex invocations. Throttles available, ordered from least invasive:
  • The wake-budget guard in dispatcher.handleEvent (default cap per minute per agent — automatic).
  • Relay-level spam filtering before the SSE event publishes.
  • For subagents that should stay internal-only, set metadata.host to a value no dispatcher matches so external mail still lands in the inbox but no worker spawns.

See also

• @agenticmail/claudecode — same pattern for Anthropic's Claude Code
• AgenticMail wiki — architecture, roadmap, operator guide
• Codex CLI — the host
• @openai/codex-sdk — the SDK we drive workers through