What Is Pulsar Agent (2026)? Cross-Harness Messaging for Coding Agents

By Antigravity Skills editorial team · Last updated

Short answer: Pulsar agent is an open-source (MIT) MCP server that puts the AI coding agents on one machine on a shared message bus. Claude Code, Codex, OpenCode, pi, omp, Cursor and Gemini CLI sessions can see each other and message one agent, a group or everyone. It runs locally; you install it with one npx command.

Not Apache Pulsar. This page is about github.com/yocontra/pulsar, a small tool for AI agents. It has nothing to do with Apache Pulsar, the distributed pub-sub messaging platform, or the Pulsar text editor. The name is shared, which is why we call this page "pulsar agent".

Skill page: the agent skills for Pulsar have their own guide, Pulsar skills: install and use.

Unofficial guide. Antigravity Skills is not affiliated with the pulsar project. Every fact below comes from the repository README, its delivery docs and the Hacker News thread, all read on November 18, 2026. The project is young and changing, so confirm commands on the official page before installing. We did not run the tool.

What is pulsar agent?

According to its README, pulsar puts "every AI coding agent on your machine on one bus". It is an MCP server, so any MCP client can use it, and it adds one machine-wide list of live agents. Each agent gets identity details on its own (harness, model, project, git branch, worktree and the pull request it works on) and can add its own tags such as role:reviewer, plus a short activity line.

You then address messages by name, by a selector such as project=apparish or model=fable, by free text, or with * for every agent. The README's example is telling everyone on one project that main is frozen until a release is tagged. The author describes it on Hacker News as a more general message bus between agents, with no opinion about how you launch them.

The repository lists the project as MIT licensed (copyright 2026, Contra), and the package is on npm as pulsar.

Which harnesses does it connect?

The README names Claude Code, Codex, OpenCode, pi, omp, Cursor and Gemini CLI, plus any MCP client. How a message gets in differs by harness. This table is our summary of the project's delivery docs:

HarnessHow a message is pushed inWakes an idle agent
Claude CodeClaude's per-session inbox socketYes
Codexcodex queue --resume through the app-server daemonYes
OpenCodeThe local server's prompt endpointYes
pi, ompThe pulsar extensionYes
Anything else (Cursor, Gemini CLI)Inbox only, with an "unread" note on each pulsar tool resultNo

For "anything else", the agent has to call inbox to see messages. The delivery docs also record limits: Codex reports no busy or idle state, so waiting for a Codex agent to go idle does not resolve, and a Codex terminal session not attached to the daemon may only show the message on resume. The docs note that an urgent mid-turn message to Claude Code was not separately verified in their testing.

How messaging works

One small daemon per machine keeps the registry of agents and a SQLite message log under ~/.local/state/pulsar. It starts on demand and stops when idle. Each agent's MCP server is a thin client that talks to the daemon over a Unix socket reachable only by your user.

The README lists these tools:

  • list_agents shows who is online, filtered by a selector.
  • send messages a name, selector or *, with a priority of urgent (mid-turn), normal (next turn, wakes idle) or fyi (no wake).
  • reply answers in the same thread, ask sends and blocks until the first reply, and wait blocks for a message, a reply or another agent going idle.
  • inbox, set_identity, whoami and history cover unread messages, tags and name, and searchable records with per-recipient receipts.

There is also a CLI (pulsar ls, send, reply, inbox, tail, doctor) for sending as yourself from a terminal.

How does it compare with Claude Code's built-in messaging?

The README includes a comparison with Claude Code's SendMessage, which we have condensed:

Claude Code SendMessagepulsar
HarnessesClaude onlyClaude, Codex, OpenCode, pi, omp, Cursor, Gemini, any MCP client
Find by project or roleNoTags and selectors
Send to a group or broadcastOne call per recipientSelector or *
Wait for a replyNoask, wait
MachinesLocal plus Remote ControlOne machine (v1)

That is the project's own comparison, not our test. For Claude Code users, pulsar's main added value is crossing harnesses and sending to groups.

How to install pulsar agent

The README says the install is one command, then a restart of your agents:

npx pulsar setup

Setup finds your agents, shows which files it will change, asks once and backs up each file it touches. It copies pulsar to ~/.local/share/pulsar/<version>, so sessions do not wait on npm at start. Useful flags are --dry-run (show the plan, change nothing), --client claude,codex (only these agents) and --uninstall. Run --dry-run first if you want to see the edits to your config before they happen.

The README also gives a curl installer for machines without Node, plus native routes: /plugin install pulsar --marketplace yocontra/pulsar for Claude Code, a Codex plugin marketplace command, and omp install pulsar or pi install npm:pulsar for those harnesses. Copy exact commands from the README, since they may change.

One detail matters for permissions: pulsar setup sets Claude Code's crossSessionInbound option to accept in ~/.claude/settings.json, so messages can reach sessions that skip permission prompts. If you manage permissions carefully, read our guide to Claude Code "Don't Ask Again" rules first. For any MCP client, the manual config is a command of npx with arguments -y pulsar mcp.

Is pulsar agent safe?

The README describes these protections:

  • The socket sits in a user-only folder, which is the trust boundary: processes running as you are trusted.
  • Delivered text is marked as coming from a peer agent, not your user, and cannot grant permissions. No agent can use the sender name operator.
  • Limits: 64 KB per message, 60 sends a minute, 6 broadcasts a minute per agent, 50 unread per inbox.
  • No telemetry. The only network call is the npm version check in pulsar doctor.

Still, agents that can instruct each other add a new way for an untrusted prompt in one session to influence another. Those are the author's claims; we did not audit the code. Treat it like any tool that edits your agent config.

Limits and what the Hacker News thread says

The project is brand new. It was posted as "Show HN" on November 17, 2026, with 21 points and 5 comments when we checked. The GitHub repo had 43 stars and its latest release at that time was 1.4.0.

  • Scope is one machine today. The README says multi-machine messaging over Tailscale is "coming soon". The author said in the thread that a next release would add it, with no cloud service.
  • A commenter compared it to herdr. The author replied that herdr is more opinionated about how you run agents, while pulsar works with any harness or terminal. See our Herdr App guide for that tool.
  • Another commenter described it as a small message broker and asked for a "done" state and redelivery. The author said they would add a fuller lifecycle (queued, delivered, seen, acked, done) and durable redelivery in a later release. That is a plan, not a shipped feature.
  • Receipts for Claude Code, Codex and OpenCode say sent, because pulsar cannot see whether the harness delivered or held the message.

Because the tool is so recent, we found no independent reviews. Expect rough edges.

How is it different from a skill?

A skill (a folder with SKILL.md) teaches one agent how to do a task. Pulsar is a running service that connects several agents. It does install a short skill that tells agents when to use its tools, so the two work together; see how to install skills for that side. Pulsar also is not a self-hosted assistant like OpenClaw, and it is not a role-based skill pack like gstack.

FAQ

Is pulsar agent the same as Apache Pulsar?

No. Apache Pulsar is a distributed pub-sub messaging platform for servers. Pulsar agent is a local MCP server for AI coding agents, in yocontra/pulsar.

Is pulsar agent free and open source?

Yes. The repository and package.json state the MIT license. You still pay for the model access your agents use.

Does it work across multiple computers?

Not yet. The README says one machine in v1, with multi-machine support over Tailscale listed as coming soon.

How do I remove it?

The README lists --uninstall for pulsar setup, which undoes what setup did.

Sources