InterAgentMail

InterAgentMail gives local Codex project agents durable mailboxes, MCP tools, and automatic wake-up delivery. Messages are stored as JSON on disk, so they remain queued while Codex or the receiving agent is offline.

One shared Codex app-server and one background InterAgentMail supervisor manage every registered project. The supervisor watches every mailbox but opens an isolated app-server connection only when that mailbox has new work, then releases it after the Codex turn is idle. Users do not need a bridge terminal, a port per agent, or copied Codex session IDs.

Requirements

• Python 3.10 or newer

• The OpenAI Codex CLI installed, available on PATH, and signed in

• Windows, macOS, or Linux; background operation is tested most heavily on Windows

Related MCP server: mcp-dispatch

Install

Install the isolated command-line application from PyPI with pipx:

pipx install cgl-interagentmail

Register one or more Codex projects, then start delivery:

iam setup "C:\Projects\MainApp" "C:\Projects\SecurityReviewer" "C:\Projects\UXReviewer"
iam start

Open the correct agent from any registered project:

cd C:\Projects\SecurityReviewer
iam open

The complete Windows, macOS, Linux, upgrade, troubleshooting, and uninstall instructions are in docs/INSTALL.md.

Browser messaging

Version 1.3 adds an authenticated browser surface backed by an ordinary human IAM mailbox. Agent replies return to the same inbox, and messages queue normally when agents or delivery services are offline.

Configure the default this-PC-only interface, choose a password when prompted, and start only its standalone companion process:

iam web setup MyMailbox --display-name "My Name"
iam web start

Open http://127.0.0.1:8787. This does not start, stop, or restart the IAM supervisor or Codex app-server. Use iam web status, iam web password, and iam web stop for its lifecycle.

LAN access is an explicit opt-in:

iam web setup MyMailbox --display-name "My Name" --lan --acknowledge-network-risk
iam web start

LAN mode uses plain HTTP plus application-password authentication; it is not end-to-end encrypted. Enable it only on a trusted, password-protected WPA2/WPA3 network, allow only Private networks in the operating-system firewall, keep the password private, and never port-forward the IAM web port. Anyone who gains access can read messages, send agent instructions, and potentially cause project changes or corruption.

Source code, releases, and issue tracking are hosted at https://github.com/cerberusgamelabs/cgl-interagentmail.

Everyday commands

iam status
iam doctor
iam report
iam user create ADDRESS
iam web setup ADDRESS
iam web start
iam web status
iam open [PROJECT]
iam restart
iam stop
iam stop --all
iam register [PROJECT ...]
iam unregister [PROJECT ...]
iam capabilities --json

• iam status shows services, projects, and pinned thread IDs; --json provides stable schema 1.0 output.

• iam doctor runs read-only health checks for IAM, Codex, services, project registration, MCP configuration, mailboxes, safety policy, and resumable threads.

• iam report creates a privacy-sanitized Markdown support report under the IAM data directory.

• iam user manages human mailboxes; iam web manages the separate authenticated browser companion.

• iam open resumes the saved project thread, or starts a new remote session when the project has not needed one yet.

• iam stop stops mail delivery but leaves the shared app-server running.

• iam stop --all stops both IAM-managed background services.

• iam unregister removes IAM's managed MCP block and project registration while preserving mailbox data.

Setup behavior

iam setup is safe to run again. It:

1. Initializes the project's durable mailbox.

2. Records its project directory and safety policy.

3. Adds a clearly marked mcp_servers.interagentmail block to .codex/config.toml.

4. Establishes the existing-inbox baseline so old mail does not unexpectedly trigger work.

5. Lets the supervisor attach the project only when undelivered mail arrives; an active or approval-paused turn retains its own connection, and idle connections are released.

Use iam setup --process-existing when existing inbox messages should be delivered immediately.

Installed releases keep data under %LOCALAPPDATA%\InterAgentMail on Windows or ~/.local/share/interagentmail on macOS/Linux. Set INTERAGENTMAIL_HOME before setup to use another shared data directory.

Project folder basenames become mailbox addresses. IAM records mailbox ownership and refuses to register two different project roots with the same address before changing either project or mailbox.

Automation and reviewer platforms

Version 1.2 adds a stable JSON interface for tools that create reviewer projects or manage agent fleets:

iam capabilities --json
iam register "C:\Projects\SecurityReviewer" --json
iam status --json
iam doctor --project "C:\Projects\SecurityReviewer" --json
iam unregister "C:\Projects\SecurityReviewer" --json

iam register is the automation-oriented alias for iam setup and is safe to repeat while delivery is running. Human-facing identity remains optional: --display-name supplies a label, but IAM does not force a persona. See docs/INTEGRATION.md for the versioned envelope, stable errors, collision behavior, identity ownership, and full lifecycle contract.

Diagnostics and support reports

Run a read-only installation check at any time:

iam doctor

Warnings describe optional or currently stopped components. Failures produce a nonzero exit status and identify configuration that needs attention.

Create a report suitable for attaching to a support issue:

iam report

The report contains software versions, operating-system information, service health, registered-project checks, mailbox counts, and bounded log statistics. It replaces project names, display names, project paths, user paths, email addresses, thread IDs, message IDs, and common credential formats. It never reads message bodies or chat contents into the report and deliberately omits raw app-server logs because those logs can contain source code or private instructions.

Use iam report --stdout to inspect or pipe the report, iam report --output PATH to choose its location, and iam report --log-lines N to change how many trailing log lines are counted. Existing output files are preserved unless --force is supplied. Automated sanitization is intentionally conservative, but review any report before sharing it publicly.

Safety

New agents use workspace-write with on-request approvals. The unattended supervisor rejects interactive approval requests instead of granting them. A task that needs approval remains uncompleted until a person opens that agent.

Only for a trusted project that genuinely requires unrestricted filesystem and network access:

iam setup "C:\Projects\SecurityReviewer" --full-access

The browser interface has application-level password authentication, CSRF protection, login throttling, and restrictive browser headers. Its LAN transport is still plain HTTP and must be treated as trusted-network-only. InterAgentMail's file-backed core is a local coordination mechanism, not an authentication boundary. Any local process that can write to its data directory can inject or modify mail. Do not share that directory with untrusted users or accept untrusted message content as instructions.

Sending mail

Agents normally use the project-bound InterAgentMail MCP tools. People and fallback workflows can use the CLI:

interagentmail send --project-root "C:\Projects\MainApp" --to SecurityReviewer --subject "Security review" --body "Review the current release and send back actionable findings."

The receiving supervisor wakes the correct Codex thread. The agent reads the message through MCP, performs the work, sends a substantive reply when appropriate, and archives the message only after it is handled.

Protocol and low-level bridge details are in SYSTEM.md. Release history is in CHANGELOG.md.

Support

InterAgentMail is free and open source. If it helps your agents work together, you can support Cerberus Game Labs on Ko-fi.

License

InterAgentMail is open-source software released under the MIT License. Use it, modify it, distribute it, and build on it. Copyright 2026 Cerberus Game Labs.

Contributions are welcome; see CONTRIBUTING.md.