How Claude Code Enables Cross‑Session AI Chat via Summarized Messages

Claude Code recently introduced a feature that lets Claude sessions on different terminals send messages to each other.

Only a summary is transmitted, not full context

The feature transmits a text summary written by Claude, not the entire conversation history or files. This avoids privacy concerns and prevents unrelated content from polluting the target session. Claude decides what information is worth sending and formats it as a clean message.

Messages are bidirectional. One session can answer another's question, and Claude can proactively notify a session when its changes affect the other.

When it is useful

The official docs list concrete scenarios:

Handoff detection : If a session detects a destructive change, Claude automatically summarizes and notifies the affected session, intercepting issues earlier than merge time.

Parallel worktree coordination : Different sessions editing the same repository in separate worktrees can inform each other of their changes, giving AI‑side support for multi‑branch development.

Long‑running task status : After a migration or test finishes, a session can report the result without the user switching contexts.

Cross‑machine reply : Sessions on other machines or web clients can reply to messages, though they cannot initiate new conversations.

Developers reported that the feature appeared within five hours of a request, confirming real demand.

How messages are delivered

The receiving Claude reads messages during tool‑call idle periods, without interrupting ongoing tasks. If the receiver is idle, Claude Code starts a new turn immediately; otherwise messages are queued and displayed as a single “Message from” line that can be expanded with Ctrl+O.

Message acceptance depends on the permission mode of both sides. The crossSessionInbound setting controls this with three options: accept (allow), hold (store), and refuse (reject).

Note that receiving a message incurs the same token‑based cost as a normal prompt.

Limitations

Only plain text can be transmitted; structured data is not supported.

Supported on macOS and Linux; native Windows is unavailable.

Platforms such as Amazon Bedrock and Google Cloud Agent Platform are not supported.

Sessions cannot approve each other's permissions or change configurations; messages are purely textual.

Getting started

Upgrade Claude Code to version v2.1.224 or later. In a session run /list-agents (or /peers) to view reachable sessions.

Each session has a default name derived from its working directory, e.g., myapp-3f. Use the /rename command to customize it.

The desktop Claude Code client on macOS and Linux also supports this feature.

Thoughts

When multiple sessions work in parallel, synchronizing information manually is costly; Claude now automates part of that.

This is not the same as an agent‑team collaboration model. Cross‑session messages act like a small pipe between independent sessions, whereas an agent team lets a single Claude manage multiple sub‑agents.

Users still decide which sessions to run and what tasks to assign, but at least the Claude instances can now converse about some matters.

Session handoff is a common problem across projects. Open‑source projects like tokenbank already implement handoff and cross‑application summaries (e.g., Claude Code to Cursor or Workbuddy). Claude Code’s integration shows the demand, though the idea is not entirely novel.

Some developers joke that their Claudes can finally chat, while others worry about being excluded from the loop; the design mitigates this because messages are only text and do not grant any approval or operational permissions.