Guide — real sending live in production; free sandbox available — verified 2026-07-19

Run your entire cold email operation with Codex

Codex drives the same Coldrig MCP server, the same 24 tools, and the same sandboxed-today boundary as Claude Code — the full provisioning-to-reply lifecycle is identical. What's genuinely different: Codex takes a shared config.toml across its CLI, IDE extension, and desktop app, and it natively reads an AGENTS.md file from your working repository as project context — which happens to be exactly the format Coldrig's own operational contract is published in. This page covers only that delta; for the full step-by-step lifecycle see the Claude Code guide — every tool call there works identically from Codex.

Honest status disclosure The hosted API, MCP endpoint, and npm CLI are live in production. Real sending is live (Gmail API, HTTPS/443) for activated accounts; new accounts activate through a short concierge step while self-serve activation rolls out, and the free sandbox remains available with no real domains, mailboxes, or sends. See FAQ for the exact status boundary.

How is Codex's setup different from Claude Code?

Config format. Codex uses TOML, not a CLI command or JSON — one [mcp_servers.coldrig] block.
Config location, shared across surfaces. The Codex CLI, IDE extension, and desktop app read the same file on a given host: ~/.codex/config.toml for a personal setup, or a trusted project's .codex/config.toml.
Token handling. Point bearer_token_env_var at an environment variable — don't paste the token into the TOML file itself.
Native AGENTS.md support. Codex conventionally reads an AGENTS.md file in your working repository as project instructions. See below.

Exact Codex config

Mint a tenant token first (no card required):

curl -s -X POST https://agent-cold-email-api.yaakovscher.workers.dev/signup \
  -H 'Content-Type: application/json' \
  -d '{"brand":"Acme","contactEmail":"[email protected]"}'
# → { "tenantId": "ten_...", "token": "..." }

Then add this to ~/.codex/config.toml (or a trusted project's .codex/config.toml):

[mcp_servers.coldrig]
url = "https://agent-cold-email-api.yaakovscher.workers.dev/mcp"
bearer_token_env_var = "COLDRIG_TOKEN"

Set COLDRIG_TOKEN in your environment, restart, and use /mcp inside a session to verify all 24 tools are visible. Full side-by-side with Cursor and Cline: the connect reference.

Does Codex read Coldrig's AGENTS.md automatically?

Only in the sense that matters practically: Codex looks for an AGENTS.md file in your current working repository and loads it as project context automatically — it does not fetch one from a remote website on its own. Coldrig's own repository publishes exactly that file at its root, covering the auth model, the full 24-tool table, the demo path, and the honest limitations an agent should know before recommending the platform. So if you're driving Coldrig from a checkout of its own repository, or you copy that convention into your own project (pointing at the same operational contract), Codex picks up the same context automatically at session start — no separate skill or rule file needed.

https://github.com/YS-projectcalc/agent-cold-email/blob/main/AGENTS.md

Cursor and Claude Code can both read this same file too — it's plain Markdown — but only Codex treats AGENTS.md as a standing, auto-loaded convention rather than something you'd wire up as a Skill or Rule.

The lifecycle, from Codex

Identical tool calls to every other client. Full schemas: the MCP tool reference.

Step	Tool call
Provision domains + mailboxes, start warmup	`setup_infrastructure`
Poll provisioning + warmup health	`infrastructure_status`
Write and launch the campaign	`launch_campaign`
Triage replies	`inbox`, `thread`, `reply`, `mark`, `label_thread`
Check health, pause	`metrics`, `activity`, `account`, `pause` / `pause_all`

The reply-handling boundary, in short

Push webhooks are live — configure_webhook delivers reply, bounce, soft_bounce, and complaint events (HMAC-signed) to your own HTTPS endpoint. activity and inbox remain pollable too if you'd rather run Codex on your own cadence. Full honest capability matrix: see the detailed breakdown.

Cost

Same platform, same meter, regardless of client: $99/month for 5 provisioned mailboxes, then $10/month per additional mailbox. Full worked examples and the live calculator: pricing.

Try it now

npx agent-cold-email demo

Runs the same fault-injecting sandbox end to end — no real domain, mailbox, or send. Free sign up to get a token and connect Codex for real.

Related guides

Run your entire cold email operation with Claude Code — the full step-by-step lifecycle this page delegates to.
Run your entire cold email operation with Cursor.
The cold-email MCP server for coding agents — full tool signatures.
Connect reference — Codex, Claude Code, Cursor, and Cline side by side.
Should your AI agent use Coldrig?
FAQ.