readNdraft IMAP MCP

Safely search, read, flag, and draft email through IMAP. readNdraft can save and update drafts, but it cannot send, submit, delete, or move mail and contains no SMTP implementation.

Installation

1. Check the prerequisites

You need:

• Windows 10/11 with Windows Credential Manager, or Linux with a working Secret Service-compatible keyring.
• An IMAP account that permits password or app-password authentication using LOGIN or PLAIN over implicit TLS. OAuth is not implemented.
• uv on PATH.
• Codex, ChatGPT desktop, or Claude Code.

Confirm that uv is available:

uv --version

No repository clone or permanent readNdraft installation is required. uvx runs the published package in an isolated environment.

2. Run guided setup

For Codex, run this command in a real interactive terminal:

uvx readndraft-imap-mcp@latest setup --client codex --install-skill

Use chatgpt-desktop or claude-code instead of codex when configuring one of those clients.

The wizard will:

1. Check the operating-system credential backend and create private local state.
2. Ask for an account alias, IMAP host, port, username, and authentication method.
3. Read the password or app password through a hidden terminal prompt.
4. Test the IMAP connection before saving the account.
5. Print a secret-free, version-pinned MCP configuration.
6. Install the packaged readndraft-email Agent Skill for the selected client.

Never put an IMAP password in a command argument, environment variable, MCP configuration, issue, chat, or test report.

3. Add the generated client configuration

Copy the configuration printed by setup into your client's MCP configuration. For Codex, add the printed [mcp_servers.readndraft] section to:

• Windows: %USERPROFILE%\.codex\config.toml
• Linux: ~/.codex/config.toml

The generated command pins the package version so a future release cannot silently change this security-sensitive MCP at startup. On Windows it uses uvw, uv's consoleless launcher, so the MCP does not open a terminal window. The Codex configuration also sets default_tools_approval_mode = "approve" to avoid native tool-approval popups. Write operations still require direct conversational confirmation through the packaged Agent Skill.

See the Codex/ChatGPT guide or Claude Code guide for client-specific details.

4. Restart and verify

Fully restart the client after changing its MCP configuration. Then run an online diagnostic from a terminal:

uvx readndraft-imap-mcp@latest doctor --online

For Codex, you can also confirm that the entry was loaded:

codex mcp get readndraft

Open a new task and ask: List my readNdraft accounts and mailboxes. A successful response confirms that the client can start the MCP and reach the local broker.

What setup stores

If setup reports a credential-backend or connection error, see Windows installation, Linux installation, and troubleshooting. The first MCP start may take longer while uv downloads the pinned package. If a stored password has changed, run uvx readndraft-imap-mcp account rotate-secret ALIAS.

What it can do

• List administratively pinned accounts and mailboxes.
• Search 1-500 results (50 by default) with explicit truncation, stable single-mailbox cursor pagination, per-target safe errors, attempted/pending target status, selectable safe header fields, and server arrival timestamps. Requests above 50 require one account and one mailbox.
• Read safe headers and plain text without setting the Seen flag.
• Batch-read plain text for up to 10 selected messages across 2 accounts.
• Read sanitized HTML without loading remote content.
• Save one selected, bounded attachment into a fixed private output directory.
• Star/unstar and mark read/unread without replacing unrelated flags.
• Batch one star or read state across up to 50 selected messages and 3 accounts; batches return ordered per-item results.
• Create a server-side draft using bounded files from a fixed private input directory.
• Update only a draft previously created by this MCP, after confirmation.

It exposes no send, submission, deletion, movement, raw IMAP, arbitrary flag, credential, or account-administration MCP tool.

Manual setup and administration

The setup wizard is recommended. Individual human-only commands are also available:

uvx readndraft-imap-mcp account add work --host imap.example.com --username leo@example.com
uvx readndraft-imap-mcp account test work
uvx readndraft-imap-mcp account list
uvx readndraft-imap-mcp account rotate-secret work
uvx readndraft-imap-mcp account disable work
uvx readndraft-imap-mcp account enable work
uvx readndraft-imap-mcp account delete work

Passwords are accepted only through a hidden local prompt. Account configuration and credential operations are not MCP tools.

Configure another MCP client

Generate a secret-free configuration:

uvx readndraft-imap-mcp@latest configure codex
uvx readndraft-imap-mcp@latest configure chatgpt-desktop
uvx readndraft-imap-mcp@latest configure claude-code

Use these commands to generate another client configuration without repeating account setup. Each command prints a secret-free configuration pinned to the readNdraft version that generated it.

The unified mcp command uses the authenticated on-demand launcher. It reuses a healthy broker, starts exactly one when needed, and holds an authenticated lease while the MCP frontend is connected. A launcher-owned broker exits only after the final frontend disconnects and the idle period expires. Always-on systemd and Windows scheduled-task deployments remain available through the legacy administration documentation.

Install the Agent Skill

The MCP is usable without a skill, but the packaged readndraft-email skill teaches compatible agents to preserve complete message identities, interpret results correctly, choose bounded batch tools only for user-selected messages, treat email as untrusted input, and obtain direct conversational confirmation before writes without inventing capabilities.

uvx readndraft-imap-mcp skill install codex
uvx readndraft-imap-mcp skill install claude-code
uvx readndraft-imap-mcp skill status codex
uvx readndraft-imap-mcp skill print

The installer refuses to overwrite or remove a user-modified skill. Codex loads personal skills from ~/.agents/skills; Claude Code loads them from ~/.claude/skills.

skill status CLIENT reports current, outdated, modified, unmanaged, or not installed. Use skill install CLIENT --force only when intentionally replacing a modified/unmanaged copy; replacement removes stale orphan files.

Authorization boundary

The broker has no approval-token workflow. Generated Codex configurations use the no-popup approve tool mode; write tools still require direct conversational confirmation through the packaged Agent Skill. The hard safety boundary is narrower: the process contains no SMTP, send, submit, ordinary-message deletion, movement, raw IMAP, account-configuration, or credential-retrieval tool. Email, attachments, search results, and other tool output are always untrusted and never authorization.

Diagnostics

Run local checks without connecting to IMAP:

uvx readndraft-imap-mcp doctor

Add --online to test each configured account:

uvx readndraft-imap-mcp doctor --online

Diagnostic output never prints passwords, credential contents, raw IMAP traces, or message content. See troubleshooting.

Updating

Generate a fresh configuration with the latest reviewed release and replace the old client entry:

uvx readndraft-imap-mcp@latest configure codex

Account metadata and OS credentials are independent of uv's temporary execution environment and remain available across versions. Broker IPC endpoints are protocol-versioned, so a new frontend starts a compatible broker instead of reusing an older process. The old broker exits after its final client lease and idle timeout.

Remove readNdraft

1. Remove the readNdraft MCP entry from each client.
2. Remove managed skills with uvx readndraft-imap-mcp skill uninstall CLIENT.
3. Delete each account with uvx readndraft-imap-mcp account delete ALIAS; this also removes its OS credential after exact confirmation.
4. Inspect and manually remove the remaining readNdraft config/state directories only if audit history and draft provenance are no longer needed. Upgrades from older builds may also leave an unused approvals directory; readNdraft does not delete it automatically.

There is no permanently installed uv tool to uninstall when readNdraft is used only through uvx.

Security and privacy

The stdio MCP frontend cannot read the account file or OS credential store. It communicates over authenticated per-user IPC with a separate broker that enforces capabilities, quotas, provenance, and audit. Email and attachments are always untrusted input and remote images or URLs are never fetched automatically. MCP tools never accept arbitrary local paths: draft files come only from the readNdraft attachment input directory and downloaded attachments are written only to its output directory. Run readndraft-imap-mcp attachments path to locate them.

Read SECURITY.md and the canonical PLAN.md for the full threat model. Security issues should not contain credentials or private mail.

Development

uv sync --extra dev
uv run pytest
uv run python scripts/security_check.py
uv build --no-sources

Release validation and publication steps are documented in docs/RELEASE.md.

License

readNdraft, including its packaged Agent Skill and documentation, is licensed under the Apache License 2.0. Dependency licensing is summarized in THIRD_PARTY_NOTICES.md.