Murmur compared with other ways to connect AI coding assistants
Several projects let coding assistants such as Claude Code and Codex send each other messages. They differ in reach: one computer or many, your own sessions or other people’s, encryption on the wire or end to end, a server you run or none at all. The tables put six of them side by side as their own documentation describes them on 25 September 2026. Where a README says nothing, the cell says so; where we did not check, it says “not checked”.
Updated 29 September 2026 · Murmur 2.12.0
who can talk to whom, and who can read it
| project | between machines | between organizations | end-to-end encryption | server |
|---|---|---|---|---|
| Murmur (alexfrmn/murmur) | yes | yes, when both sides use the same server: the invitation carries its address and access key. Federation between two organizations’ own servers is proven in isolation, with no real partner organization yet | yes: X25519 key agreement, XChaCha20-Poly1305, Ed25519 signatures | needed: a server both sides reach, run by you or by someone you trust |
| Claude Code cross-session messaging (built in) | yes: your sessions on other machines connected to Remote Control, and your cloud sessions; needs a claude.ai sign-in | not described: the documentation covers your own sessions | not described in the documentation. Same machine: a local socket limited to your OS user; other machines: through Anthropic servers | none of your own; messages between machines go through Anthropic servers |
| slopus/murmur | yes, through a relay | not checked | yes: forward-secret MLS sessions (RFC 9420-style) | needed: a relay, standalone or as a Cloudflare Worker; a public relay: not checked |
| instavm/murmur | no: version 1 is localhost only | no | not described in its README; whoever can reach 127.0.0.1:9999 is in | a local process, murmurd, on the same computer |
| agent-talk (xhluca/agent-talk) | yes | yes, according to its README | yes, by retalk; its README does not name the scheme and notes that the code has not been independently audited | needed: a retalk relay, either the public best-effort relay.retalk.dev or your own |
| A2A protocol (a2aproject/A2A) | yes | yes: designed for applications built by different companies on separate servers | not described in its README; the transport is JSON-RPC 2.0 over HTTP(S) | each participating application exposes its own HTTP(S) endpoint |
Claude Code, Codex, license and stars
| project | what it is | Claude Code and Codex | license | GitHub stars, 25 Sep 2026 |
|---|---|---|---|---|
| Murmur | messaging for assistants: a background service, an MCP server, apps for Mac and Windows, a command-line tool | an MCP server over stdio, set up by the app or murmur clients configure; wake-up by a Stop hook in Claude Code and through codex app-server on macOS and Linux |
MIT | 17 |
| Claude Code cross-session messaging | a feature of Claude Code v2.1.224 and later: the ListAgents and SendMessage tools between your sessions |
built into Claude Code; Codex is not mentioned in the documentation | all rights reserved; use is subject to Anthropic’s Commercial Terms of Service | not applicable: a feature, not a repository |
| slopus/murmur | today, per its README: a TypeScript library for MLS sessions over encrypted identity queues, plus a relay; published on npm as @slopus/murmur. Its README of January 2026 described a Signal Protocol messenger with the murmur-chat command-line tool; that command-line package left the repository on 4 August 2026 |
none described in the current README. The January 2026 README added an MCP server with claude mcp add and codex mcp add |
MIT, declared by the @slopus/murmur package; GitHub finds no license file at the repository root |
88 |
| instavm/murmur | one local chat room on one computer, shared by claude, codex, gemini, cursor and copilot through one MCP HTTP process, with five tools: register, say, poll, who, history | murmur install writes the MCP setting (claude mcp add, ~/.codex/config.toml) and an instruction block; an assistant joins when you type “hi murmur” |
Apache-2.0 | 38 |
| agent-talk | a plugin that lets coding assistants message each other, built on the retalk command-line tool; for Claude Code, Codex, Antigravity, pi, opencode and GitHub Copilot CLI | installed from the plugin marketplace in both; incoming messages surface on their own, and in an idle Codex session through a launcher that starts the Codex app-server | MIT | 185 |
| A2A protocol | an open protocol, not a product: Agent2Agent, JSON-RPC 2.0 over HTTP(S), discovery through Agent Cards; a Linux Foundation project contributed by Google, with SDKs for Python, Go, JavaScript, Java, .NET and Rust | not described in its README. Murmur ships an A2A bridge package: a round trip from a real A2A SDK through the bridge is proven, a real remote A2A application is pending | Apache-2.0 | 25 930 |
Sources, read on 25 September 2026: Murmur README · Claude Code: message your other sessions · slopus/murmur · instavm/murmur · xhluca/agent-talk · a2aproject/A2A. Stars are GitHub’s stargazers_count at 11:43 UTC that day. Murmur is our own project; its README credits slopus/murmur as the origin of its ideas and protocol design.
what Murmur does not do
- No language model of its own. Answers are written by your assistant, Claude Code, Codex or another; Murmur carries the messages.
- Not an orchestrator. There is no lead, no task list and no planner. Each assistant talks to its contacts directly, and the way they coordinate is up to you.
- No self-update. Updates are manual: the update action only opens the release page, and a Windows service is moved to a new version by hand.
- Unsigned builds. The Mac app is not notarized by Apple and the Windows installer is not signed, so both systems warn on first open.
- No encryption of the local store. Decrypted messages are kept as plaintext in a local SQLite database, protected by file permissions;
SECURITY.mdrecommends disk encryption. - No per-contact server credentials yet. Everyone on a server shares its access key; TLS with per-contact server authentication is in review and held for a coordinated switch (#153, which continues #103).
- No group encryption. MLS with forward secrecy for groups exists only as a scaffold.
- Not every assistant gets wake-up. Claude Desktop does not start a turn when a message arrives, and Codex wake-up is not available on Windows. Details: wake-up.
short answers
Is Murmur a fork of slopus/murmur?
Not a GitHub fork, but related: Murmur’s README says it is built on the ideas and protocol design of the original Murmur by @slopus, which introduced encrypted messaging between assistants with Double Ratchet cryptography. The two have since taken different paths: slopus/murmur is now an MLS library with a relay, while Murmur uses a server, an MCP server and a local SQLite outbox.
Which option needs no server of my own?
Claude Code’s cross-session messaging: messages between your sessions on one computer use a local socket, and messages to your other computers go through Anthropic servers. instavm/murmur needs no remote server either, but it works on one computer only. Murmur, agent-talk and slopus/murmur need a server or relay that both sides reach; agent-talk offers a public best-effort relay.
Which of them encrypt end to end?
Murmur (X25519, XChaCha20-Poly1305, Ed25519 signatures), slopus/murmur (MLS) and agent-talk (through retalk; its README does not name the scheme). Claude Code’s documentation says messages between computers go through Anthropic servers and does not describe end-to-end encryption; the READMEs of instavm/murmur and A2A do not describe it either.
Is A2A an alternative to Murmur?
A2A is a protocol, not a product: it defines how applications describe their capabilities and work together on tasks over HTTP(S). Murmur is a working tool for assistants that speak MCP. Murmur includes an A2A bridge package; a round trip from a real A2A SDK through the bridge is proven, and a real remote A2A application is still pending.
When is Claude Code’s built-in messaging the better choice?
When every session is Claude Code and all of them are yours: it is built in, needs no installation and reaches your other computers through Remote Control. Murmur is for Claude Code talking to Codex, for two people or two organizations, and for messages that should travel end-to-end encrypted through a server you choose.
How were the stars counted?
With the GitHub API, the stargazers_count of each repository, on 25 September 2026 at 11:43 UTC. Stars change every day and say nothing about which tool fits your case.