SIAgentNews.com
← All news

Pulsar lets AI coding agents from different vendors message each other

A new open-source project puts agents like Claude Code, Codex and Gemini CLI on one local message bus. It is early, but it shows where agent tooling is heading.

· First reported by github.com

64 KBsize cap on a single Pulsar message

The short version

  • Pulsar connects agents such as Claude Code, Codex, OpenCode, pi, omp, Cursor and Gemini CLI on one machine.
  • Agents can message one agent, a group chosen by tags, or every agent, and a message can wake an idle one.
  • Messages are labeled as coming from a peer, never the user, and cannot approve anything.
  • Limits include 64 KB per message and 50 unread messages per inbox; the project says it has no telemetry.
  • It is MIT-licensed and installs with one command, but it is a very young project.

Why it matters

Anyone running several AI coding agents today often copies results between them by hand. A neutral message layer could remove that chore. It also creates a new route for bad instructions to reach an agent, so the safety design matters as much as the convenience.

What Pulsar does

Pulsar is a project posted to Show HN, the Hacker News section where makers show off their work. As github.com reported in the repository listing, it puts every AI coding agent on your machine onto one shared message bus. The agents it names are Claude Code, Codex, OpenCode, pi, omp, Cursor and Gemini CLI.

Once connected, an agent can see which other agents are running, label itself by what it is working on, and send messages. A message can go to one agent, to a group, or to everyone. Notably, a message can wake an agent that is sitting idle.

The project's own example shows a list of three agents in one project, each with a tool name, a code branch and a status of busy or idle. The user then tells all three that the main branch is frozen until a release is tagged, and the tool reports how it reached each one.

How it is put together

Pulsar is an MCP server, which is a standard way to plug extra tools into an AI agent. Any agent that supports MCP can use it. For agents without a way to be pushed a message, messages wait in an inbox, and every Pulsar tool result says when unread ones are waiting.

Behind the scenes, one small background program (a daemon) runs per machine. It starts when needed and stops when idle. It keeps the list of live agents and a log of messages in a SQLite file, which is a small database stored as a normal file. Each agent talks to the daemon through a local channel that only your user account can reach.

The daemon delivers each message using whatever route the agent's own tool supports. It also scans for Claude Code and OpenCode sessions that never loaded Pulsar, so those show up in the list and can receive messages too.

  • Addressing: by name, by a tag filter such as a project or branch, or to every agent.
  • Tags: project, tool and branch are added automatically, and agents can add their own, like a workstream.
  • Replies: agents can reply in a thread, ask and block until an answer arrives, or wait for a message instead of looping on a timer.

Safety rules and their limits

Letting agents talk to each other raises an obvious worry: one agent's text becomes another agent's input. Pulsar's answer is to mark every delivered message as coming from a peer agent, not from you, and to state that it cannot grant permissions. Each agent can accept, hold or refuse incoming messages. No agent can use the sender name reserved for the human operator.

There are also caps. A message can be up to 64 KB, an agent can send 60 messages a minute and 6 broadcasts a minute, and an inbox holds 50 unread messages. These stop two agents from replying to each other endlessly. The project says it collects no telemetry and makes no network calls except a version check in its diagnostic command.

The trust boundary is simple. Anything running as your user account is trusted, the same as it is with your files. That is reasonable for a local tool, but it means the protection against malicious messages rests largely on the receiving agent's own guardrails. As our hosts noted, an inbound agent message is just text entering a model's context, so a human check should stand between messages and risky actions like deleting files or spending money.

Installing it, and how early it is

Setup is one command, npx pulsar setup. It finds your agents, shows which files it will change, asks once, and backs up every file it touches. It copies itself to a local folder so agent startup does not wait on the network. Rerunning the latest version upgrades it in place. A separate script installs a private copy of Node for people who lack it, and the code is released under the MIT license.

There are also command-line tools for people: list who is online, send a message, read replies sent to you, watch a live feed of joins and messages, and run a health check.

Still, this is a young project. At the time of our briefing it had one point and no comments on Hacker News, so there is no track record and no outside review. The idea is mostly incremental, since message queues are decades old. What is new is the neutral inbox that agents from competing vendors can share.

Who should care

The audience is mainly developers who run more than one coding agent at once. A typical workflow would have one agent writing code, another reviewing it and a third running tests, with no copy-pasting in between.

For everyone else, the story is a preview. As agents multiply, the tools that connect them may matter as much as the agents themselves. Non-developers can safely skip installing it.

What to watch

  • Will other agent tools adopt a shared messaging approach, or will each vendor build its own?
  • How well do the peer-message labels and rate limits hold up against prompt injection, where hidden instructions are slipped into text an agent reads?
  • Does the project gain users, outside security review and steady updates?

Try it

If you run several coding agents, try Pulsar in a throwaway project folder with no secrets. Read the list of files setup says it will change before you approve, and keep a human review step on anything an agent does from a message.

Sources and links

First reported by github.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news