All stories
Pidgeon Ecosystem Community reports

Pidgeon gives AI agents an inbox, and humans the last word

The agent-email service's October 29 updates add enforced human takeover, document reading and more MCP tools, alongside a new Show HN presentation.

Sources (6)

Why now Pidgeon's October 29 changelog documents new handoff enforcement and tools; its same-day Show HN post introduces the service to a fresh developer audience without establishing its first launch date.

Pidgeon wants an AI agent to have its own email address, while leaving a person able to take over the conversation. Its October 29 updates make that handoff an API restriction: according to the changelog, an API key attempting to reply, forward or send a draft in a thread taken over by a person receives 403 taken_over.

The key also cannot simply switch the conversation back to software. A person hands it back from the inbox; the agent can still prepare a draft for review. That makes the update useful for support, booking and follow-up workflows where an agent and an operator might otherwise both answer the same customer.

The service appeared in an October 29 Show HN post. It already has earlier dated releases, so today’s presentation is not evidence that the entire product first launched today.

Email tools for an existing agent

Pidgeon is a hosted email service, not a new model or a fork of Codex. Its API and TypeScript and Python SDKs expose addresses, messages and threads. A hosted MCP server connects those resources to existing agents.

The October 29 changelog says the MCP server and AI SDK integration add 13 tools covering calendar events and invitations, forwarding and draft management. It also adds text extraction for PDF, Word and spreadsheet attachments up to 10 MB, with the extracted content screened for hidden instructions. These are documented provider capabilities, not an AITrending integration test.

For Codex, the documented route is to add the hosted MCP endpoint with codex mcp add pidgeon --url https://pidgeon.ai/api/mcp, then sign in using codex mcp login pidgeon. The guide also offers a plugin bundling the server and skill, or an environment-held PIDGEON_API_KEY for unattended use. Access can be limited to an address and permission level; read-and-draft access is designed to let an agent prepare messages without sending them.

What the protection does, and does not, establish

Pidgeon’s Firewall documentation describes rules for hidden text, fake chat roles, instructions addressed to a model and requests for secrets. It says the verdict is produced by rules without sending the mail to an AI model. Cleaned text is the default for its MCP and AI SDK tools.

Suspicious-mail holding became the default for new workspaces on October 27. Existing workspaces retain their previous behavior until an owner enables it. At an address with holding enabled, the company says suspicious content stays unreadable to API keys and app passwords until a person releases it. A clean verdict means no rule fired, not that an email is safe to obey; the service also says it does not scan attachments for malware.

At review time, the Free plan lists three inboxes, 3,000 emails a month and a 100-email daily limit. Developer is listed at EUR 20 a month for ten inboxes and 10,000 monthly emails, with paid add-ons. Pidgeon says it hosts the service in the EU and does not train on mail content.

An early HN commenter asked how shared-domain sending reputation is protected from another customer’s outreach. That question illustrates the operational issues beyond connecting a tool. Today’s documented change is narrower and concrete: once a person takes over a thread, the agent’s key is supposed to stop sending into it.