Claude Code v2.1.224: Cross-Session Messaging, Self-Hosted Environments & Auto Mode Explained (2026)

Claude Code v2.1.224, released the first week of August 2026, adds three major changes: sessions can now message each other via ListAgents and SendMessage, self-hosted environments let Team/Enterprise orgs run cloud sessions on their own infrastructure, and auto mode becomes the default permission mode for new CLI sessions starting August 14. This guide covers what each feature does, how to set it up, and what it means if you're running Claude Code from outside the US.

Cross-session messaging: your sessions can now talk to each other

Before v2.1.224, running multiple Claude Code sessions in parallel meant manually copy-pasting context between terminals — telling the frontend session what the backend session just changed, or pasting an API contract from one window into another. Cross-session messaging removes that step.

Two new tools power it:

• ListAgents — discovers other active Claude Code sessions on the same machine (and, as of v2.1.225, named Remote Control sessions on other machines too).
• SendMessage — sends a text message to another session, either because you asked Claude to, or autonomously when a change in one session affects what another is working on.

A message is plain text Claude writes for the other session — it never includes your conversation history or files, so there's no accidental context leak between projects. The receiving session shows a Message from row once Claude reads it; press Ctrl+O to expand it. Run /list-agents anytime to see which sessions are reachable.

A typical workflow looks like this:

# In the session working on the frontend
> Tell the session working on the payments API that users.name is now users.display_name

The other session receives the message, reads it, and can reply. This is the foundation for what Anthropic calls Agent Teams — a lead session coordinating multiple worker sessions (frontend, backend, tests) that stay in sync without you relaying messages by hand.

Availability: macOS and Linux only for now, on v2.1.224+. Windows support hasn't landed yet.

Security note: if a session is running with bypassed permissions, inbound cross-session messages are held for your approval by default (the crossSessionInbound setting). Also as of v2.1.222, outbound SendMessage calls now go through the permission classifier before dispatch, so a session in auto mode can't silently spam other sessions.

Self-hosted environments: run cloud sessions on your own infrastructure

Also shipped in early August: self-hosted environments, in public beta for Team and Enterprise plans. Instead of cloud sessions (started from claude.ai, the mobile/desktop apps, or claude --cloud) always running on Anthropic's infrastructure, you can now point them at your own machines or containers.

Setup, as an Owner or admin:

# 1. Turn on "Allow self-hosted environments" in admin settings
# 2. On your own machine/container:
claude self-hosted-runner setup

The guided setup creates the environment and starts a runner. Once it registers, it shows Healthy in admin settings, and anyone on the org can pick it when starting a cloud session — from any surface (web, mobile, desktop, CLI). Runners come in two modes: fixed (a set number always running, load-balanced across them) or on-demand (an orchestrator spins runners up and down to match queue depth).

This matters for regulated teams (finance, healthcare, public sector) who need Claude Code to touch internal databases and services without exposing them publicly, and for compliance requirements where code and data shouldn't leave the org's own network. Note this is distinct from Remote Control, which lets you continue a session already running on your own laptop from your phone — self-hosted environments run on shared infrastructure your platform team operates, usable by any authorized user.

It's off by default and not available on ZDR (zero data retention) plans, and Anthropic explicitly recommends a platform/DevEx team own the runner image and ongoing operation if you turn it on.

Auto mode becomes the CLI default

Starting August 14, 2026, auto mode is the default permission mode for new Claude Code CLI sessions on Pro, Max, and Team plans. If you'd already set a personal default mode, it's preserved unless you accept a one-time switch prompt; an org-managed default also isn't overridden. You can switch modes at any time regardless.

To lock in auto mode as your default ahead of the switch (or to set it explicitly), add this to your settings:

{
  "permissions": {
    "defaultMode": "auto"
  }
}

One practical upside already in effect on those plans: classifier calls that auto mode makes to decide whether an action needs approval no longer count against your usage limits, so the safety layer itself doesn't eat into your quota.

Other changes bundled into v2.1.220–v2.1.224 worth knowing: the subagent spawn cap was removed, a new archive plugin source lets you install plugins from a zip over HTTPS without git or npm (with optional SHA-256 pinning), Bash commands can no longer hide part of themselves from permission checks via tab or invisible-Unicode padding, and the Ultraplan research preview (including the /ultraplan command) was removed in favor of plan mode or Claude Code on the web.

Running Claude Code with a third-party API endpoint

None of these features change how Claude Code authenticates to the model — it still needs a working ANTHROPIC_BASE_URL and API key, whether you're on the official Anthropic API or a relay. That matters more now, not less: with self-hosted environments and multi-session workflows, you're likely running more concurrent sessions than before, which means more API spend and more exposure to rate limits and regional access issues.

This is where a service like Safa API fits in. It's a unified relay that speaks the same OpenAI-compatible / Anthropic-compatible interface Claude Code already expects — you just swap the base URL and key. It's noticeably cheaper than official per-token pricing, supports Claude's prompt caching so repeated context across sessions doesn't get re-billed at full price, doesn't require a US credit card (Alipay works), and routes Claude, GPT, and Gemini through the same endpoint so a multi-session Agent Team setup can mix models without juggling three separate accounts. For developers outside the US who don't want their cross-session Agent Team workflow to stall on a single provider's access restrictions, that single-endpoint setup is a practical hedge.

Frequently Asked Questions

Do I need to upgrade to use cross-session messaging?

Yes — you need Claude Code v2.1.224 or later, and it currently only works on macOS and Linux. Run claude update or reinstall to get the latest version, then use /list-agents to confirm other sessions are visible.

Will auto mode being default make Claude Code less safe?

Auto mode still goes through a permission classifier before taking risky actions, and as of v2.1.222 that classifier also screens outbound cross-session messages. If you prefer stricter control, you can set defaultMode to something else in your settings and it won't be overridden by the platform default.

Is self-hosted environments the same as running Claude Code fully offline?

No. Self-hosted environments still connect to Anthropic's model API — only the compute running the session (file access, shell execution, internal network reach) moves to your infrastructure. You still need a working API endpoint, whether that's the official Anthropic API or a compatible relay.