Of all the agents we have wired into our outreach stack, ChatGPT is the one clients ask about first and the one with the most fiddly answer. Not because the tools are different — they are the same tools — but because whether you can attach a custom connector at all depends on which ChatGPT plan you are on.
So this article is mostly about the boundaries. What an agent can do with your outreach, what it cannot, and which door you go through depending on what you pay for.
The plan question, answered properly
WarmySender’s ChatGPT and Codex guide lays the plans out, and it is worth reading them exactly rather than assuming:
- Free — no. Custom connectors are not available.
- Plus — developer mode is listed, but full connector support for write actions is not there. In practice this means the Codex route below is your path.
- Pro — yes, in developer mode, read-only.
- Business — yes; an owner turns on developer mode for the workspace.
- Enterprise / Edu — yes; an admin grants developer access.
On Business and Enterprise, write actions require a confirmation before they execute. We consider that a feature and have not tried to switch it off.
The practical upshot for most agencies: if you are on Plus or Pro and want an agent that can actually build and launch things rather than just read, go through Codex.
The Codex route
Two lines in a terminal, no key:
codex mcp add warmysender --url https://warmysender.com/mcp
codex mcp login warmysenderThe second opens a browser sign-in against your own WarmySender workspace. WarmySender’s ChatGPT cold email page puts the setup at about a minute, with no API keys required.
If you are running headless — a shared box, a scheduled job, anything without a browser to click in — there is a key path instead:
export WARMYSENDER_API_KEY=ws_YOUR_API_KEY_HERE
codex mcp add warmysender --url https://warmysender.com/mcp --bearer-token-env-var WARMYSENDER_API_KEYKeys are created in Settings → API Keys and can be scoped read-only if you want an agent that looks but does not touch. We use read-only keys for anything that generates reports.
For the packaged version with skills bundled on top, the plugin repo is public at github.com/fortuneflick/warmysender-claude-plugin, installable in any agent that supports skills with npx skills add fortuneflick/warmysender-claude-plugin.
What the agent can do
WarmySender’s server page describes 78 tools in four groups: around thirty read tools, roughly fifteen write tools for email and prospects, about twenty for LinkedIn, Instagram and warmup, and eight for email verification.
For cold email specifically, the agent can create a multi-step campaign, edit it, and set the sending window and daily limits. It can start, pause and resume. It can add prospects in bulk, work from your saved templates, and manage the suppression list. On warmup, the Codex warmup page shows it can turn warmup on for a connected mailbox and choose its strategy and target volume, apply the same settings across a whole set of mailboxes in one instruction, and list mailboxes with their health before you scale sending.
That “one instruction across twelve mailboxes” case is the one that earns its keep for us. Onboarding a client with a dozen fresh mailboxes used to be a dozen identical trips through a settings screen.
What the agent cannot do
This is the part we put in every client-facing explanation, because it is the part that determines whether this is safe to use on someone else’s revenue.
It never sends. “ChatGPT never sends a message itself and can never raise a limit.” When it creates or launches a campaign, it writes the campaign and hands it to WarmySender’s scheduler. The scheduler paces every email, LinkedIn, Instagram and WhatsApp action inside safe limits, and email verification shares the same allowance and pacing.
It cannot raise a limit. Not with a persuasive prompt, not with a clever argument, not by accident. Daily caps, domain ceilings and the gradual ramp are enforced by the scheduler regardless of who initiates the campaign. There is simply no tool that turns them up.
It cannot connect or disconnect a mailbox. Account connection stays in the app. Mailbox credentials never pass through an agent or a chat log.
It cannot delete. The available tools cover pause and archive, not deletion.
It cannot reach another workspace. Each key or sign-in reaches only its own workspace’s data.
How we actually use it day to day
Three patterns, in descending order of how often we reach for them.
Reporting, constantly
“Which campaigns sent nothing yesterday and why.” “Which mailboxes look unhealthy.” “How is the fintech sequence doing against the SaaS one.” These are read-only, they are safe on any plan that supports connectors, and they replaced a genuinely tedious daily ritual.
Bulk setup at kickoff
New client, new mailboxes, new lists. Verify the list, set warmup across every mailbox with the right strategy, build the campaign shell, load the suppression list from the client’s existing customers. Then a human opens the app and reads all of it before anything runs.
Triage when something looks off
“Explain why this campaign is quiet.” Nine times out of ten the answer is a limit or a window doing exactly what it should, and knowing that in ten seconds instead of ten minutes is the whole value.
Our rule for clients
An agent may build. A human launches after reading. We have not found a case where skipping the read saved enough time to be worth it, and the failure mode of a wrong-but-well-formed campaign is a burned list, which is not recoverable in a week.
FAQ
Can I add a custom connector on ChatGPT Free?
No. Free plans cannot add custom connectors.
What about Plus?
Developer mode is listed for Plus, but full connector support for write actions is not available. Use the Codex route, which is included with Plus and above.
What does Pro give me?
Custom connectors in developer mode, read-only.
How do Business and Enterprise differ?
On Business, an owner turns on developer mode for the workspace. On Enterprise or Edu, an admin grants developer access. Write actions require a confirmation before they execute.
Can the agent raise my sending limit if I insist?
No. There is no tool for it, on any plan.
Does connecting an agent change how fast my email goes out?
No. The scheduler paces sending identically whether a human or an agent created the campaign.